This fixes many of the Windows+GPU crashes.

Issues #96, #100, and more...

This saves 4 bytes from LightBVHNode (which doesn't matter, since it goes
from 32 to 28 bytes) and potentially gives more convergent node access in
the PDF() method.

This gets LightBVHNode down to 32 bytes. On the GPU, this gives a 5-10%
performance improvement in the material / BSDF eval kernels.

Along with some further packing, this gets it down to 8 bytes +
sizeof(LightBounds), (vs 28 bytes before.)