Add more stress tests and benchmarks

[talex5]

The only stress test we have is stress_semaphore.ml and the benchmarks mostly test the synchronisation primitives but not the OS integration (except for bench_fd.ml, which tests reading from /dev/zero). We need more.

In particular, we have found several resource leaks and races when e.g. spawning and joining domains in a tight loop (e.g. #421). The https://github.com/ocaml-multicore/retro-httpaf-bench/tree/master/cohttp-eio benchmark is useful; more like that (or simpler) would be good.

Testing for FD leaks would also be worth doing. e.g. checking that we have the same number of open FDs at the end of running the tests.

It would be very interesting to look at the benchmarks with e.g. https://magic-trace.org/ to see if there are obvious inefficiencies left.

The benchmarks could go in this repository (if simple) or we could create a new project just for that (if they have lots of dependencies).

Stress testing https://github.com/ocaml-multicore/lwt_eio would be useful too.