Previously, we'd add up contributions for all the vertices of a path and
then divide by wavengths' PDFs in Film::AddSample()/AddSplat(). The problem
with this was the case where secondary wavelengths are terminated in the
middle of path (e.g. for dispersion). What we actually should be doing (and
now are) is dividing by the current wavelength PDFs any time we add to the
radiance sum for a pth.  Thus, vertices before the secondary wavelengths
are terminated aren't weighted incorrectly, by the post-termination PDFs...

Related to issue #23.

Issue #20.

Work-around to get the GPU working in release mode (though performance will
be for now non-optimal..)

Issue #20...

Remove sensor "exposuretime" parameter. Instead pass through time serived from camera shutter open/close times.

e.g. sm 6 GPUS (10 series) on Windows...

(This came up in issue #21).

Add a comment that one of the argument of `cudaMemAdvise()` is unused,
to avoid the reader wondering why `0` is being passed instead of
`deviceIndex`.

`cudaMemAdvise()` and `cudaMemPrefetchAsync()` on GPU devices require
those devices to have the `cudaDevAttrConcurrentManagedAccess` attribute
set. Since pre-fetching is a performance optimisation and is not
mandatory, only enable it when the attribute is set and skip it
otherwise to avoid errors (see [this comment][0]).

[0]: https://github.com/mmp/pbrt-v4/issues/20#issuecomment-683182824

Any file which is not part of a target will not show up in Visual
Studio, regardless of whether it is mentioned by a `source_group()` or
not.

Noted in issue #20.

Fixes warnings that were noted in issue #20.

Mostly to make MSVC happy, but it's got a point...
Relates to issue #20.

Fixes via @wuyakuma in issue #20 (with some reformatting / minor changes.)

This helps a lot with high-resolution volumes...

This led to NaNs in highly scattering media...

The previous version was buggy in that by looping over leaves, it didn't
consider active voxels at higher levels of the tree.

It also seems to have been hurting performance by using the leaf's maximum
value to set the density. Considering density pointwise in the nanovdb grid
ends up giving an overall 1.37x speedup for bunny-cloud and 2.80x for
explosion(!!)

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