Update shared hashmap to use RwLock instead of threadlocal

[mattcompiles]

Motivation

While doing some performance testing I found that the threadlocal version of the resolver cache was have a large negative impact on build times. Reverting back to the RwLock version is almost halving the asset graph build times in native builds. I have also validated that V2 builds are also performing better with this change, however not by the same margin.

Changes

• Migrate back to RwLock Hash maps for the resolver cache

Checklist

• Existing or new tests cover this change

[yamadapc]

This doesn't revert the entire PR that moved to thread-local ; I'd rather revert everything or at least understand where this makes sense vs. not at the moment.

I have also validated that V2 builds are also performing better with this change

How much better are V2 builds? It's weird because this did not cause a regression when rolled-out on our metrics.

It could be that this strategy works for the caches you didn't revert but not the packages/tsconfigs case or that it works for none & something is wrong with benchmarks/measurements (for example the CI agents we're using have a lot more memory than your computer).

[mattcompiles]

This doesn't revert the entire PR that moved to thread-local ; I'd rather revert everything or at least understand where this makes sense vs. not at the moment.

I did originally go to revert but the revert commit needed changes anyway so I decided to just do this file as it was what I had experimented with.

How much better are V2 builds? It's weird because this did not cause a regression when rolled-out on our metrics.

I only did one test just to validate it wasn't a regression for V2 and I built the asset graph about 20s quicker, 4:48 for with thread local vs 4:25 with this change.