Optionally allow as much computation with CUDA as possible

### Description Since PyTorch has the ability to perform tensor operations on a CUDA device, and zuko is built on PyTorch, there is no reason why a zuko flow could not be trained on GPU. This may provide a substantial performance increase when accelerator hardware is available. ### Implementation Most tensor operations can be carried out with CUDA already. Exceptions may include integer matrix multiply, which does not appear to be supported by PyTorch. Any operation that cannot be performed on GPU should be done on CPU and then migrated if needed. ### Alternatives No alternatives considered at this time.