BUG: `is_array_api_obj(np.matrix(...)) == True`

[lucascolley]

(Pdb) A.A
matrix([[1, 2, 3],
        [4, 5, 6]], dtype=int32)
(Pdb) scipy._external.array_api_compat.is_array_api_obj(A.A)
True

[lucascolley]

also

(Pdb) scipy._external.array_api_compat.is_numpy_array(A.A)
True

it looks like we are deliberately accepting np.ndarray subclasses, but I think we should special case rejecting np.matrix as a known bad-case. WDYT @ev-br ?

[ev-br]

Ouch. Reject all ndarray subclasses, I'd say? Maybe , just maybe, whitelist np.memmap. actually not even that, not until someone checks it, someone who uses it (scikit-learn?). Other than memmap, just plain ndarray /whatever is needed for numpy scalars , np.generic.

[lucascolley]

could you send a PR or would you prefer to review?