Fixed an incorrect radiance conversion factor in 'coating' and 'roughcoating' reported by Attila Afra.

The models now produce the right falloff compared to Monte Carlo
reference renderings with explicitly instantiated layers

Implementation of the paper 'Accurate computation of single scattering in participating media with refractive boundaries' contributed by Nicolas Holzschuch

- fixed several bugs that occasionally broke double-precision builds
- rewrote min/max binning code to make it tolerant to precision issues.
  Previously, it gave up in such cases, which lead to a slow build
  with the O(n log n) method.

texture formats
matrix 2x2 inversion special case (wasn't called before)
more meaningful flags on vertex.