[llvm] Reland "[lit] Migrate lit to ProcessPoolExecutor (#202681)" (PR #209076)

[Cullen Rhodes via llvm-commits]

c-rhodes wrote:

> @c-rhodes I've opened a "potential" fix for this #214386. It replaces the `wait()` based collection loop with `add_done_callback` on each future instead of rescanning the whole pending set every time, so it should remove the cost that scales with `LIT_SUBMISSION_WINDOW`. I can't reproduce the original regression on my hardware, so I can't confirm it end to end. Would you be able to rerun the `LIT_SUBMISSION_WINDOW` values you tried before against that branch and see if the slope flattens out?

this brings the testing time down to 50s for all `LIT_SUBMISSION_WINDOW` values, so still ~8% worse but a big improvement, thanks for the quick fix

https://github.com/llvm/llvm-project/pull/209076