[Profile-Guided Optimization][Feature] Let practitioners specify a trade-off between speed and memory when selecting backends for FX graphs

Nobody has claimed this yet.

fyi @kiya00

Currently our profile-guided optimization attempts to pick a backend for each FX graph by looking at speed or memory use. We should consider letting practitioners specify a function of speed and memory to decide on a backend. This would essentially be letting practitioners customize the generic compilation function with a mechanism to select a backend based on the statistics the auto-benchmarking gathered.

A further refinement of this idea is that we should make the backend selection global -- considering all FX graphs at the same time -- and update our default compilation to select the fastest backends s.t. we don't OOM when storing grads for every FX graph.

We should consider these refinements once we have the optimization of FX graphs in isolation with a focus on speed working as expected.

cc @borda

Start by locating the profile-guided optimization and FX graph backend-selection entry points, then review how auto-benchmarking statistics currently support speed- or memory-based choices. Clarify the intended speed/memory trade-off API and whether global selection is in scope; done should include a defined selection mechanism that works for isolated FX graph optimization before the proposed global refinement.

Written by the indexing model from the issue text.