This speeds up light BVH builds by ~20%.

The lights now all use an InternCache for their DenselySampledSpectrum
values, saving a significant amount of memory for scenes with many
area lights that have the same emission spectra. (Partial fix to issue #183.)

This way things like seg faults will no longer cause pbrt to just
silently terminate, which has been fairly confusing behavior...

- Use a std::atomic for the bytesUsed statistic in the BufferCache (a legit
 bug, though harmless.)
- Hold associated mutexes when consuming futures. Strictly speaking, all of
  the worker threads should be idle when these are consumed and it's only the
  main thread that's running, so there should be no need...

Fixes #179, via @shadeops.

sigma was removed in f5ad66a5

Assorted renamings in the VolPathIntegrator implementation.

Partially seeding an RNG using just ray direction is risky business with an
orthographic camera. Things go extra bad when the object enclosing the
medium has camera-axis aligned planes in which case doing the rest of the
RNG seeding with tMax to the intersection point rounds out the borkage.

Fixes #175.

Fixes #176.

s/BumpEvalContext/NormalBumpEvalContext/
Factor Bump() into BumpMap() and NormalMap()

Switch to DiffuseBxDF in DiffuseMaterial
Remove "sigma" parameters with --upgrade

- Adds chi^2 tests for various microfacet model configurations
  (conductive, dielectric, anisotropic, isotropic, dense->less dense and
  vice versa, etc.)

- Increased the directional resolution of the chi^2 test

This commit fixes issues in the computation of Jacobian of the rough
dielectric half-direction mapping and the criterion for discarding
back-facing microfacet normals with respect to both the incident and
outgoing direction argument. These changes now enable the rough
dielectric to pass the chi^2 tests.

Rename texture mapping classes to not have "2D" and "3D" suffixes.
Otherwise just naming and formatting changes.

pZownZ's x and y components are 0, so....

We need to transform dpdx and dpdy to texture space first... (Not doing that
could lead to overblurring or aliasing, depending on the circumstances.)

Just variable renaming and formatting...

This takes a mutex (that should be held at call time). It returns the
future's value if it's available and otherwise unlocks the mutex, tries to
do some work, and then returns with the mutex re-locked.  This allows
mutex-protected future harvesting to work without deadlock...