Rebase of #610 and also move emscripten up much higher in the hierarchy to
ensure that it doesn't have too much of a ripple effect on other platforms.

This involved moving down a good number of definitions, but hopefully was done
with care to not break anything!

This works by specifying a "runner" for actually executing the binary.
This doesn't apply to the Android or NetBSD runs because there
isn't a simple binary that just runs the executable.

We now create an additional binary `linux_fcntl` for testing this
since there are header conflicts when including all necessary headers.
This binary is run on all platforms even though it's empty on all non-
Android/non-Linux platforms.

Testing has been switched from a custom binary to using a runner-less
test (or pair of tests). This means that for local development a simple
`cd libc-test && cargo test` will run all the tests. CI has also been
updated here to reflect that.

It's now broken due to changes in the `gcc` crate and having a too-old compiler,
and in general it's unfortunately architecturally so different from the other
test frameworks that it's difficult to maintain over time.

We can't test s390x because qemu segfaults but we can at least verify that it
compiles.

Closes #650

It’s not clear if these values are used or needed by the simulator. The doc only seems to mention arm values for the actual device:
https://developer.apple.com/library/content/documentation/General/Reference/InfoPlistKeyReference/Articles/iPhoneOSKeys.html#//apple_ref/doc/uid/TP40009252-SW3

Pass the tests only if we find an expected positive output (rather than checking for negative output to fail the tests).

Documentation generation appears to be failing on master (https://travis-ci.org/rust-lang/libc/jobs/183483333):

```
Collecting ghp-import
  Downloading ghp-import-0.4.1.tar.gz
Collecting travis
  Could not find a version that satisfies the requirement travis (from versions: )
No matching distribution found for travis
```

Basically `--user` doesn't appear to take an option, so `pip install ghp_import --user $USER` makes pip think the user is a package that needs to be installed (in this case `travis`). As there is no `travis` package, it dies.

This helps caching between runs and can help speed up turnaround time for
various operations. The old android container didn't work out for some reason
due to permissions so the definition has now been vendored locally to just
rebuild it each time.

Also cache on appveyor to speed up builds.

* Override the compiler via CC; ./configure can't seem to detect it.
* Unset CROSS_COMPILE when running make. Per the comment, we otherwise
  end up invoking commands like i686-ar.

Specifically:

* Update the base docker images to ubuntu 16.10
* Update musl to version 1.1.15

This is necessary since the versions we were previously using do not
define CLONE_NEWCGROUP, so adding the symbol to the libc crate causes
test failures.