ENH: average to allow better use w/ custom objects

[numpynewb]

A simple syntax change to enable better use of average with custom Python objects.

This works because if weights is None, then avg.dtype can be "O" when working with custom Python objects.

The fix does nothing but change how np.shape is called, so that it will work with objects that don't necessarily have shapes, like base Python floats, custom Python objects, etc.

[tylerjereddy]

If this is useful, my first thought is that it would be helpful to have a unit test that demonstrates something that works with this change that didn't work before. That would also provide a sort of "code example" for the behavior you're trying to enable.

[tylerjereddy]

The macos Azure failure is probably not related -- we may also have to skip / decorate the test in numpy.f2py.tests.test_semicolon_split.TestCallstatement.test_callstatementas I think I saw that stochastically fail once before as well.

[numpynewb]

@tylerjereddy thank you for your quick response, and apologies for the delay. This is actually not as simple as I had initially thought when submitting the change, because, for example, a.mean(axis) can generate a base-Python float, which will also return an error in some instances where a dtype is needed. Bear with me while I put together a coherent vignette (I am somewhat new to contributing, so apologies for what I now understand to be an incomplete commit/merge request!).

I believe that this is important (selfishly, of course) because it will facilitate the use of average, and other NumPy routines that rely on it, with custom numeric-like Python objects, where arbitrary arithmetic and NumPy operations make sense.

Thanks again; I very much appreciate you responding.

[charris]

Needs release note.

[seberg]

Going to close the PR, I expect its still relevant. But right now this is an incomplete one-line change with tests missing and was idle for many years.

If anyone wants to pick it up, np.mean for example does work-around the same issues pretty successfully, so there should be something to copy. I think gh-8696 tracks the issue also.