Use atomic flag for buffer memory-attribute initialization

[seberg]

Store and read Buffer._mem_attrs_inited via std::atomic using acquire/release semantics to avoid races during lazy pointer memory- attribute initialization under free-threaded execution.

At least strictly speaking, _mem_attrs_inited is a flag that guards access to other memory attributes.
My understanding is that on ARM processors with weak memory ordering this can be a problem even beyond compiler optimizations making use of the fact that order isn't strictly guaranteed (without the light use of atomics).

I.e.:
thread1 -> init _mem_attrs
thread1 -> sets _mem_attrs_inited
thread2 -> reads _mem_attrs_inited
thread2 -> reads _mem_attrs

Does not guarantee that thread2 doesn't read the uninitialized _mem_attrs

This is probably not sufficiently covered by pytest-run-parallel, so if desired I/we should add an explicitly threaded test for it. (Although I have doubts we'll be able to fail the test even without the changes)

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[seberg]

/ok to test 7c1c215

[leofang]

The change seems fine to me. Though I wonder if there'd be a significant perf hit. Originally Buffer was meant to be a thin wrapper over the raw pointer, but we've grown the complexity over time.

Might be a good idea checking with @danielfrg and see if the micro benchmarks can tell us the perf difference. I dunno if/where it runs TBH.

In any case, approving. Thanks, Sebastian!

[seberg]

if there'd be a significant perf hit

I doubt this'll even register, I think the critical sections may be measurable but still so fast that it shouldn't really matter.
(I don't think atomics have much of an impact on non-threaded code, and even if threaded this should scale well, I believe.)

[github-actions]

Doc Preview CI
Preview removed because the pull request was closed or merged.