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

TODO

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

The issue names no files or tests. Start by locating the image-embedding, prompting, and mask-generation service entry points, then run benchmarks on V100s for the first and subsequent inferences with torch.compile. Done means recording whether subsequent n=1 inference achieves the proposed 20% speedup and comparing first-inference cost.

Written by the indexing model from the issue text.