The value computed in the presence of chromatic media was slightly incorrect..

Now, it returns a SampledSpectrum representing Tmaj, either from [0,tMax)
or from the last sample position returned via a callback if one or more
samples was taken.  In turn, this value is used to correctly update various
path throughputs and sampling PDFs in the integrators.

This fixes some subtle bugs from mishandling that before. As a bonus, with
this change, volumetrics on the GPU are now about 2x faster, though it's
not obvious why that would be (and I haven't dug in further yet.)

And in turn, stop grabbing values out of the rayQueue; all queues are now
only accessed by the kernels that consume them.

The precomputed sample values have moved to PixelSampleState, for now at least.

No longer take an instance to world transform; instead, apply that
transformation in the caller when needed (OptiX). Beyond being cleaner,
fixes a rare case before where we'd intersect a small non-degenerate
triangle, transform its vertices in InteractionFromIntersection() and end
up with a degenerate triangle...

Issue #12...

This lets the Windows GPU build run efficiently, without the hack of
synchronizing after each kernel. There, the GPUPathIntegrator is now stored
in CPU memory; in turn, the GPU kernels capture *this by value so that they
can access the values of members.

This led to NaNs in highly scattering media...

Cloudy-bunny is over 2x faster; its shadow ray kernel is over 5x faster
(with a small increase in noise.)

Resolve MixMaterials immediately in the closest hit shader. Among other
things, this fixes bump mapping with them. Also, issue error messages for
MixMaterial cases not handled on the GPU path.