fourier_feature_nets

Addresses some parts of the code that were causing problems with newer versions of Pytorch.

Various improvements in this release:

1. New trex_400 and trex_800 datasets, perfect for testing view-dependent effects, self-occlusion, and thin structures.
2. cuda device flag is now configurable everywhere. I don't recommend running this on CPU, but the scripts/notebook no longer require a GPU
3. The volume_raycasting animation is much cleaner and easier to follow now
4. OcTree algorithms are a bit cleaner
5. RayDataset handles the Sparse sampling mode correctly now
6. Miscellaneous bug fixes

First version of the library.