chore: Reduce contention when using EszipV2 from multiple threads

[andreubotella]

The EszipV2 struct guards the module map behind an Arc<Mutex>. This is needed to support streaming sources while returning a result before finishing parsing. However, this means every operation that needs access to the module map has to contend for that same lock.

This contention is inevitable while the streaming parsing is going on, but reads also contend with each other. This change avoids this read contention by switching the Mutex for a RwLock.

However, for pending modules, the get_module_source and get_module_source_map methods add a waker for the current async context to the module's waker list, and this would need some write locking. To avoid this, this change also guards the waker list behind a Mutex, which also reduces the scope of that contention by making it specific to a module.

With these changes, given a fully-resolved EszipV2 which isn't modified, all operations on it are now guaranteed to be lock-free.

[andreubotella]

PTAL @lucacasonato @dsherret

[lucacasonato]

I don't know if this is actually a net positive - do you have benchmarks?

[andreubotella]

I put together a benchmark based on denoland/deno#17663, which is basically a stress test on loading a realistic module graph (the Fresh server). It's at https://github.com/andreubotella/deno-compile-worker-benchmark. Let's close this PR I guess.

With Mutex:

cpu: AMD Ryzen 7 3700X 8-Core Processor
runtime: deno 1.30.3 (x86_64-unknown-linux-gnu)

file:///home/abotella/Projects/forks/denoland/deno-compile-worker-benchmark/bench.ts
benchmark                 time (avg)             (min … max)       p75       p99      p995
------------------------------------------------------------ -----------------------------
parallel stress test  616.04 ms/iter  (609.5 ms … 625.98 ms) 618.57 ms 625.98 ms 625.98 ms
single test            99.64 ms/iter   (97.5 ms … 102.98 ms) 100.12 ms 102.98 ms 102.98 ms

With RwLock:

benchmark                 time (avg)             (min … max)       p75       p99      p995
------------------------------------------------------------ -----------------------------
parallel stress test   619.7 ms/iter (611.36 ms … 625.71 ms) 624.04 ms 625.71 ms 625.71 ms
single test           101.23 ms/iter  (98.92 ms … 103.71 ms) 101.82 ms 103.71 ms 103.71 ms