closes #28883

Here's an attempt to use the incoming OpenBLAS feature allowing to register a user threading callback. The goal is to make OpenBLAS use scikit-learn's OpenMP threadpool instead of its own. See the linked issue for more details about the benefits.

The PR is not ready for merge:

• this feature will be available in OpenBLAS 0.3.28 which is not released yet. I'm testing it locally for now.
• this feature has issues that may or may not be solvable because of the way we handle flexible OpenMP threadpool sizes (see Thread callback for OpenMP backend OpenMathLib/OpenBLAS#4770)
• there's no way to make this callback use a dynamically defined threadpool size.
• it has a side effect on the whole program, not just within scikit-learn, because it register the callback on all linked openblas libs, which can be used by other libraries. It would be possible to make a context manager that unregister the callback on exit. I don't see us doing that everywhere though.
• registering the callback at import time as done in this PR for now should be avoided because of the previous point and because it requires to instantiate the threadpool controller at import time which we tried and caused a deadlock in 1.5.0.