Hacky workaround for bug #151.

This also allows re-enanabling error messages for unused texture paramters
with the GPU rendering path...

In spite of the error bounds, we seem to be rarely seeing some incorrect
reintersections without this.

It's all handled in the OptiXAggregate now, without the overhead of
creating intermediate Curve shapes.

In the previous implementation, vignetting would affect the values of the
geometric quantities stored in the gbuffer, which is clearly not what
anyone wanted...

Fixes #149.

Fixes #150.

Fixes denoise-optix on Windows.

Issue #138.

(Which seems to be a strange MSVC + CUDA 11.1 interaction...)

This finally brings curves to the GPU rendering path, though presumably in
a much less efficient way than (TODO) using OptiX curves directly would.

The logic was all kind of borked, with issues including the fact that the
number of primitives in each BLP mesh was reported to be the number of
vertices.  Further, the AABBs weren't being properly initialized, etc.  It
seems to have previously worked for single patch meshes by luck, but was
completely broken for meshes with multiple patches.

Fixes #141.

Previously the build errored out with a pre-version 11 version of CUDA.
Now it, more helpfully, issues a warning and builds a non-GPU version
of pbrt.

For BxDFs where both layers are perfectly specular (e.g. coated conductor
with both interface and the conductor being perfectly smooth), we need
to return a BxDFSample that indicates that it is perfectly specular so that
integrators don't try to use MIS...

(Allows us to handle UTF-8 filenames for that as well.)

This saves the need to take 1-v in the lookup.

Switched over to the in-memory versions of the lodepng
encode/decode functions so that we can use our own UTF-8 aware file
reading/writing functions.

...and use it to issue warnings if any such images are used for
light sources.