Implement decompression with CUDA to support ML training

[nfrechette]

Machine training often uses raw or near raw animation data for training. While fast to sample, it consumes a huge amount of memory which can make it difficult for the training set to fit entirely into GPU memory. By allowing animation data to be kept compressed in GPU memory, it should be possible to fit the entire training set.

A simple PyTorch plugin/module could be introduced to handle compression and decompression as needed.