Thread limit introspection, part 2: Empirical observation in CI

[itamarst]

Fixes #214

In theory one could, say, distribute another package in PyPI/conda-forge of compiled OpenMP code so users can run this locally. Or I guess ... pick some part of e.g. scikit-learn that is unlikely to change but that seems brittle too. So this could be better, and maybe sometime it's worth doing that.

But at least it should run in CI and give us useful info, so it's a good first step.

[itamarst]

Looks like occasionally the output is erroneously "not sure", but mostly gets everything right.

[ogrisel]

Is this PR still needed if we merge #228?

[itamarst]

I think so - the technique I use for testing in #228 is Linux-specific (strace). There's apparently dtrace on macOS but that's restricted by System Integrity Protection, not sure what you'd do on Windows, etc.. So this PR is a nice OS-agnostic approach that is still informative.

[ogrisel]

Ok, I'll add it to my review backlog.

For Windows, feel free to take over or inspiration from this vibecoded proof of concept: #221.

[ogrisel]

Some improvement suggestions but otherwise LGTM.

[ogrisel]

Note: I noticed that threadpoolctl.threadpool_info fails to detect opennblas (or any other BLAS) on some Windows builds. The conda package list still reports openblas as installed and therefore probably used by numpy. This is not a problem for that PR but this is probably a new bug.

EDIT: I opened #231 to track that issue.

[itamarst]

OK I've addressed all review comments.

[ogrisel]

Merged! Thanks again @itamarst.