ENH: Convert flatiter, broadcast and nditer to heap types - #32641

Ready for review.

I don't think we can merge this, since it will break any subclassing of these types with static types.

Forgive my ignorance Matti but I was just curious, does numpy subclass those with static types somewhere or are you thinking users might?

Users might, since we export them in the API struct.

So would like to know which direction is acceptable.

1. Adding note in this release that static subclassing of these types is unsupported, then convert them a release later.

2. Heap types only under Py_LIMITED_API, static types otherwise. I know other folks are not fan of #ifdef Py_LIMITED_API as it means two definitions per type but it's confined to the type definition block and normal builds are untouched. This can go along with the release note from point 1 as well and then in later release we can drop non limited one entirely?

If neither works, then _multiarray_umath can't be built against the Limited API at all, since a static PyTypeObject doesn't compile once Py_LIMITED_API is defined.

I think we should discuss this at a community/triage meeting. In the meantime, can you scan popular github projects for code subclassing these types as static (or for that matter, even heap) types?

I searched for this and for the types in this PR no project sets tp_base to PyArrayIter_Type, PyArrayMultiIter_Type, NpyIter_Type or PyUFunc_Type. The ones mentioning them almost all are vendored NumPy headers, NumPy forks or CPython itself.

We discussed this at the community meeting and the general opinion was positive, the probability that anyone is subclassing these is quite small. I will give this one more review pass.

Other than @ngoldbaum's valid analysis on the performance cost of the change to heap types on #32552, I didn't find any problems here. One thing it might be nice to check is that the assignments of the type in PyArray_API succeed. A good test would be to add to _multiarray_tests.c.src a check that only known slots are NULL, since I don't think we have anything today that does a check. There are some slots that should be NULL, so the test would have to take that into consideration.