ENH: spatial: serialize concurrent calls to QHull

[andfoy]

Reference issue

Closes gh-8856

What does this implement/fix?

This PR introduces a Mutex mechanism to serialize QHull calls, thus preventing potential segfaults or undefined behaviour, since it is not thread-safe. See http://www.qhull.org/html/qh-code.htm#reentrant

Additional information

Depends on #20611

[ngoldbaum]

Too bad about the code churn but putting the lock cleanup in a finally block does make sense.

[ngoldbaum]

Probably worth benchmarking singlethreaded performance since acquiring an uncontested PyThread_type_lock does have some overhead.

[rgommers]

Great, looks like this fixes a long-standing crash in LinearNDInterpolator! The one failure is memory allocation in the 32-bit job:

numpy.core._exceptions._ArrayMemoryError: Unable to allocate 1.49 GiB for an array with shape (6250000, 32) and data type float64

Could you skip this test on 32-bit platforms and when not enough memory is available @andfoy? Search the code base for _IS_32BIT and check_free_memory to see the patterns for how.

Also, could you add @pytest.mark.slow? This test takes a while to run, so we don't need to do so in every CI job.

[andfoy]

Thanks for the pointers! @ngoldbaum @rgommers @ev-br, I'll make sure to fix the tests and some minor lock acquisition details

[rgommers]

The diff with whitespace hidden looks quite clean now. Probably about ready to go once the merge conflict is resolved?