BUG: interpolate: fix high memory usage for 2 classes

[Matryoskas]

Reference issue

Closes #20357

What does this implement/fix?

When using the LinearNDInterpolator and CloughTocher2DInterpolator classes, instead of keeping the data about the location and simplicies of all the input points in the memory at once, only keep the data about a single simplex at once, greatly reducing the memory usage with larger inputs

[Matryoskas]

In MacOS, psutil seems to be returning wildly inconsistent values for the memory used in the test I made: -210 Mb and 141 Mb. For context, in my Linux machine, I always get between 40 and 60 MB. I have decided to increase the tolerance to 512 MB which should still showcase the benefits of this bug fix - before the fix, in my machine, this test would try to allocate an array of size 90 GB.

[ev-br]

As a quick note, I believe it's better to add an ASV memory benchmark (https://github.com/scipy/scipy/tree/main/benchmarks) than a unit test.

[Matryoskas]

Done. Did not realize that ASV also had memory benchmarks

[Matryoskas]

I am getting an error in the Linux - 32 bit check in scipy/_lib/tests/test__util.py:364: in test_basic. I don't think this is an error on my end since this test has no dependencies from my changes. Since the last 2 days of commits to the main branch have skipped the Linux - 32 bit check, is it possible to run this check in the main branch?

[lucascolley]

Yes, that is showing on other PRs too

[Matryoskas]

@ev-br, when you have the chance, can you take a look at my pull request?

[ev-br]

Sorry about the delay @Matryoskas .

This PR looks reasonable. I checked that the main interface with qhull in interpnd.pyx is basically equivalent to what it was before #18376 which caused the memory regression.

I think we should merge this for 1.14 when the CI comes back green.

[ev-br]

... and CI is green, and so is this PR. Thanks @Matryoskas for the patch and @Kai-Striega for the review.