Revisit need for our own ufunc type resolver

[mhvk]

Currently, we override the standard numpy type resolver, because, as seen in this extract:

pyerfa/erfa/ufunc.c.templ

Lines 535 to 545 in e33ee55

	* We provide our own type resolver, since the default one,
	* PyUFunc_DefaultTypeResolver from
	* numpy/core/src/umath/ufunc_type_resolution.c, has problems:
	* 1. It only looks for userloops if any of the operands have a user
	* type, which does not work if the inputs are normal and no explicit
	* output is given (see https://github.com/numpy/numpy/issues/11109).
	* 2. It only allows "safe" casting of inputs, which annoyingly prevents
	* passing in a python int for int32 input.
	* The resolver below solves both, and speeds up the process by
	* explicitly assuming that a ufunc has only one function built in,
	* either a regular one or a userloop (for structured dtype).

However, as noted in numpy/numpy#11109 (comment), things should now work for us. So, this is a reminder to try removing our work-around.

[mhvk]

A quick test shows that one cannot just remove the resolver (numpy 1.24.2 or numpy 1.25.dev). Likely, as noted in the comment, we'd need to define our own promotors.

See numpy/numpy#11109 (comment)