XLA Autotune utility

Right now any autotuning that happens interferes with the memory high water mark. It would also be better if we could just do the autotuning ahead of time for every model once and get it out of the way, using a memory configuration that is least likely to go OoM.

Start by locating the existing XLA autotuning path and the handling of the memory high-water mark. Determine how autotuning currently affects memory and how models are identified for one-time tuning. Done means autotuning can be performed ahead of serving with a memory configuration that minimizes out-of-memory risk and no longer interferes with the high-water mark.

Written by the indexing model from the issue text.