fix(compio-runtime): correct sched affinity mask on multi cores

[YangchenYe323]

The current implementation of bind_to_cpu_set is incorrect because the underlying syscalls called by core_affinity::set_for_current (sched_setaffinity) replaces the entire affinity mask, so passing in more than 1 cores results in the thread pinned to the last core instead of the whole set. I'm not sure core_affinity crate is still maintained (last update 8 years and author no activity since 2024), so I implemented the logic in compio directly.

Disclaimer that claude wrote the code and it's not tested on non-Linux platforms, but I cross-referenced the code in core_affinity crate.

[Berrysoft]

Thanks! But compio is a single-threaded runtime. Why do you want to bind multiple cores?

[YangchenYe323]

Thanks! But compio is a single-threaded runtime. Why do you want to bind multiple cores?

For example to pin a thread to a NUMA node like this pattern

I found it's not always desirable to model everything "thread-per-core", for example, if I have a 1000 node cluste, every node has 128 cores, and runs an MPI style communication grid, naively doing the thread-per-core means every node will end up having 1000 * 128 connections to each other, which is not good. So I use a numa-aware thread pool to control that connection fanout

[Berrysoft]

Also consider using pthread_setaffinity_np for NetBSD? core_affinity also uses it.