Add clear loop implementation

[ngoldbaum]

Note that this can't be merged until numpy/numpy#23262 is merged and there's a wheel available to install.

There are a couple other minor cleanups as well, but mostly this implements custom get_clear_loop and clear_loop implementations, allowing us to call free on the embedded references stored in the numpy array before the array is deallocated.

I haven't checked, but this will hopefully fix all memory leaks, although there may be other places we need to call free as well. In any case, it makes it much easier to check for those places!

[seberg]

Looks good, besides those two comments.

[seberg]

Suggested change

	*flags = NPY_METH_REQUIRES_PYAPI | NPY_METH_NO_FLOATINGPOINT_ERRORS;
	*flags = NPY_METH_NO_FLOATINGPOINT_ERRORS;

You don't require the python API actually if you are u sing free() to free/allocate.

NumPy currently actually ignores the float flag also, because how will clearing cause that?

[ngoldbaum]

Oh, I thought NPY_METH_REQUIRES_PYAPI was required for the clear loop, I'll try without.

[seberg]

For all numpy ones, yes :).

[ngoldbaum]

To avoid a seg fault in the sorting routine I needed to add the NPY_NEEDS_PYAPI flag. There's no reason why we can't release the GIL, but to get this to work we'd need to add support in numpy for dtypes that hold references that aren't python objects (e.g. make it so NPY_ITEM_IS_POINTER can be set without NPY_ITEM_REFCOUNT while still getting data NULLing at init and the ability to call the clear function).

[ngoldbaum]

This PR needs to be merged to fix the tests, but I'll hold off on self-merging just in case Sebastian or Peyton wants to weigh in on the last commit I just pushed.

[seberg]

Meh... the numpy sorting code is being stupid there, but I guess you can workaround it here for the time being.

[ngoldbaum]

Opened numpy/numpy#23292 and reverted the last commit.

[seberg]

I retriggered the NumPy "nightlies", so you can hopefully put this in without tests failing (unless it depended on that NULL one).