Some performance improvements

[JCBrouwer]

Some progress on #4

• Refactor optimal_textures function into OptimalTextures module (this makes it easier to use PyTorch's JIT and compile APIs)
• Enable batch support
• Update defaults and add some performance related arguments to CLI:

  --no_tf32             Disable tf32 format (probably slower).
  --cudnn_benchmark     Enable CUDNN benchmarking (probably slower unless doing a high number of iterations).
  --compile             Use PyTorch 2.0 compile function to optimize the model.
  --script              Use PyTorch JIT script function to optimize the model.
  --device DEVICE       Which device to run on.
  --memory_format {contiguous,channels_last}
                        Which memory format to use for optimization.

On my 1080 Ti this version is almost twice as fast for a batch size of 1. Utilization is definitely still low, but upping the batch size helps a lot and I think this is a good springboard for realizing more gains.