Same as 64-bit divistion, except quad wide divison leads
  to inaccurate multiplier in certain cases when the log2 of
  the divisor overflow to zero.

  The multiplier factor that used for integer division wasn't correctly
  calculated due inaccurate implementation of the emulated version
  of intrinsic `_udiv128`.

NpyIter_Deallocate returns 0 for failure and 1 for success,
so the check `NpyIter_Deallocate(iter) < 0` is not correct.

(See relevant comments in gh-20069.)

MAINT: Refactor reductions to use NEP 43 style dispatching/promotion

* ENH: Create an experimental export of the new DType API

This expposes an experimental DType and ufunc API.
It can be used by including the experimental header AND
setting the `NUMPY_EXPERIMENTAL_DTYPE_API=1` environment variable.

A user must further call:

    if (import_experimental_dtype_api(version) < 0) {
        return NULL;
    }

with a strict version check that will error on any mismatch.

At this time, I _know_ that some changes will happen, but it feels like
it may be a good idea to make this public anyway, with all the problems
attached...  (e.g. as of now ufunc reductions will just not work, since
that PR is not yet merged.)

Co-authored-by: Ross Barnowski <rossbar@berkeley.edu>

Co-authored-by: Sebastian Berg <sebastian@sipsolutions.net>

ENH: Add annotations for `np.matrix`