illumos epoll_event struct should be packed

While attempting to run `cargo test` within the mio crate on illumos I noticed a number of the tests fail. Digging into the various epoll calls I discovered that the epoll_event struct was misaligned due to extra padding. The fix is to pack the epoll_event struct with the same compiler setting that the linux variant is using.

A simple rust example that demonstrates the issue:
```rust
extern crate libc;
use libc::epoll_event;
use std::mem;
fn main() {
    println!("{}", mem::size_of::<u64>());
    println!("{}", mem::size_of::<epoll_event>());
}
```
Running the above code on Linux results in:
```
8
12
```
while on illumos it currently results in:
```
8
16
```
Looking at the `test_close_on_drop` test from mio I traced the `epoll_ctl`
calls and saw the following:
```
[root@rustdev ~/src/mio]# dtrace -wn 'pid$target::epoll_ctl:entry {this->ev = arg3; printf("%d\n", arg2); print((struct epoll_event *)this->ev); stop()}' -c "/root/src/mio/target/debug/deps/test-109e1422fb40f621 test_close_on_drop"
dtrace: description 'pid$target::epoll_ctl:entry ' matched 1 probe
dtrace: allowing destructive actions
running 1 test
CPU     ID                    FUNCTION:NAME
  6  92874                  epoll_ctl:entry 4
struct epoll_event * 0xfffffc7fee7feda8
test test_close_on_drop::test_close_on_drop ... test test_close_on_drop::test_close_on_drop has been running for over 60 seconds
  6  92874                  epoll_ctl:entry 6
struct epoll_event * 0xfffffc7fee7fee18
^[[A  6  92874                  epoll_ctl:entry 7
struct epoll_event * 0xfffffc7fee7fee18
```
I dumped each of the epoll_event's with mdb:
```
[root@rustdev ~/src/mio]# mdb -Fp 219856
Loading modules: [ libumem.so.1 libc.so.1 ]
> 0xfffffc7fee7feda8::print
mdb: no symbol information for 0xfffffc7fee7feda8: no symbol corresponds to address
> 0xfffffc7fee7feda8::print struct epoll_event
{
    events = 0x80000001
    data = {
        ptr = 0xfffffffffffffc7f
        fd = 0xfffffc7f
        u32 = 0xfffffc7f
        u64 = 0xfffffffffffffc7f
    }
}
>
[root@rustdev ~/src/mio]# prun 219856
[root@rustdev ~/src/mio]# mdb -Fp 219856
Loading modules: [ libumem.so.1 libc.so.1 ]
> 0xfffffc7fee7fee18::print struct epoll_event
{
    events = 0x80000001
    data = {
        ptr = 0
        fd = 0
        u32 = 0
        u64 = 0
    }
}
>
[root@rustdev ~/src/mio]# prun 219856
[root@rustdev ~/src/mio]# mdb -Fp 2198
Loading modules: [ libumem.so.1 libc.so.1 ]
> 0xfffffc7fee7fee18::print struct epoll_event
{
    events = 0x80000004
    data = {
        ptr = 0x100000000
        fd = 0
        u32 = 0
        u64 = 0x100000000
    }
}
```
The output from the last two `epoll_event`'s represent `Token(0)` the Client and `Token(1)` from the mio test. The first one however is from `AWAKEN` which is defined as `usize::MAX`.  This value should be 18446744073709551615. However if we convert the hex value we see something else:
```
> 0xfffffffffffffc7f=E
                18446744073709550719
```
Because of the extra 4 bytes of padding currently present in the illumos `epoll_event` definition the low order bits are picking up some junk from other memory.

All of the poll-related mio tests pass with this change. Two other tests are still failing, which appears to be caused by an OS bug, not a problem with mio or libc.

Add libc definitions for HermitCore (https://hermitcore.org)

HermitCore is based on lwIP, newlib, and pthread-embedded.
Some definitions are similar to other targets using newlib, however some are different enough to justify an own "hermit" port and not base on the existing "newlib" port.

Fix OSX builders on CI

Looks like Travis has moved on from our old images, so we're forced to
update.

Add the FIO* consts for *-apple-*

Previously only FIONREAD was implemented for x86_64 apple. The constant is shared between i{3..6}86 and x86_64, so I've moved it into the shared module rather than the 64 bit specific one.

They're defined like this in sys/filio.h:
```
#define FIOCLEX	     _IO('f', 1)        /* set close on exec on fd */
#define FIONCLEX     _IO('f', 2)        /* remove close on exec */
#define FIONREAD    _IOR('f', 127, int) /* get # bytes to read */
#define FIONBIO     _IOW('f', 126, int) /* set/clear non-blocking i/o */
#define FIOASYNC    _IOW('f', 125, int) /* set/clear async i/o */
#define FIOSETOWN   _IOW('f', 124, int) /* set owner */
#define FIOGETOWN   _IOR('f', 123, int) /* get owner */
#define FIODTYPE    _IOR('f', 122, int) /* get d_type */
```

Rather than decipher the C macros I got the values like so:
```
#include <stdio.h>
#include <sys/ioctl.h>

int main() {
	printf("pub const FIOCLEX: ::c_uint = 0x%x;\n", FIOCLEX);
	printf("pub const FIONCLEX: ::c_uint = 0x%x;\n", FIONCLEX);
	printf("pub const FIONREAD: ::c_ulong = 0x%lx;\n", FIONREAD);
	printf("pub const FIONBIO: ::c_ulong = 0x%lx;\n", FIONBIO);
	printf("pub const FIOASYNC: ::c_ulong = 0x%lx;\n", FIOASYNC);
	printf("pub const FIOSETOWN: ::c_ulong = 0x%lx;\n", FIOSETOWN);
	printf("pub const FIOGETOWN: ::c_ulong = 0x%lx;\n", FIOGETOWN);
	printf("pub const FIODTYPE: ::c_ulong = 0x%lx;\n", FIODTYPE);
}

$ gcc --target=i686-apple-darwin -o x x.c && file ./x && ./x
./x: Mach-O executable i386
pub const FIOCLEX: ::c_uint = 0x20006601;
pub const FIONCLEX: ::c_uint = 0x20006602;
pub const FIONREAD: ::c_ulong = 0x4004667f;
pub const FIONBIO: ::c_ulong = 0x8004667e;
pub const FIOASYNC: ::c_ulong = 0x8004667d;
pub const FIOSETOWN: ::c_ulong = 0x8004667c;
pub const FIOGETOWN: ::c_ulong = 0x4004667b;
pub const FIODTYPE: ::c_ulong = 0x4004667a;

$ gcc --target=x86_64-apple-darwin -o x x.c && file ./x && ./x
./x: Mach-O 64-bit executable x86_64
pub const FIOCLEX: ::c_uint = 0x20006601;
pub const FIONCLEX: ::c_uint = 0x20006602;
pub const FIONREAD: ::c_ulong = 0x4004667f;
pub const FIONBIO: ::c_ulong = 0x8004667e;
pub const FIOASYNC: ::c_ulong = 0x8004667d;
pub const FIOSETOWN: ::c_ulong = 0x8004667c;
pub const FIOGETOWN: ::c_ulong = 0x4004667b;
pub const FIODTYPE: ::c_ulong = 0x4004667a;
```

I'm just awaiting an XCode install to check they're the same on arm.

Add alignment feature and use #[repr(align(x))]

Trying to solve #1042.

Here I introduce the discussed feature that will allow going from struct alignment with a private `__align` field to using `#[repr(align(x))]`. However, I have not implemented it for all structs that require alignment yet, only `in6_addr`. This because I did not want to spend too much time before we have discussed and solved the remaining questions regarding this.

One thing to discuss is testing. I have so far not done anything to the CI scripts. So currently they will still test the crate only with the `align` feature disabled. Thus they will make sure the `__align` fields are still correct. But no automatic tests make sure everything is correct when the `align` feature is turned on. What do we want to do about that? Can we insert another `cargo test` with `--features align` to make all the Travis jobs run the test suite twice, or will that slow things down too much?

I have tried using this version of libc in rustc and the standard library. And successfully changed `Ipv6Addr::new` to not use any `unsafe` and to become a `const fn`. Whether or not we want that is out of scope for this PR, but my point was that the changes introduced with this PR allow much more flexible usage of the libc structs that have alignment.

HermitCore is based on lwIP, newlib, and pthread-embedded.

Add some posix_spawn apple extension flags

Revert "Link to libgcc when statically linking musl"

This causes linker errors due to duplicated symbols. See https://github.com/rust-lang/rust/pull/52157

This reverts commit 920cfeac.

Add MODULE_INIT_ constants

These are flags required to implement the linux kernel loading mechanism.
Specifically finit_module.

These are flags required to implement the linux kernel loading mechanism.
Specifically finit_module.

Implement statfs for dragonfly, freebsd and openbsd

Openbsd fix

put in several fixes for OpenBSD (one fix per commit).
testsuite ran on OpenBSD 6.3-current (upcoming 6.4)

```
RUNNING ALL TESTS
PASSED 6474 tests
```

they doesn't exists on OpenBSD. put them under NetBSD specific tree.

Fix netbsd build with "stdbuild"

A bug was introduced in #930 which caused netbsd to not link to libm.

r? @alexcrichton