torch.compile for faster n=1 batch runtime after the first inference

[rbavery]

TODO

• test first inference time with torch.compile
• test time for subsequent inferences on V100s
• see if we get a 20% speedup