NumPy returns Python integers in that particular case, and I am not sure I feel it's worth changing it.

The reason is, that if a place only uses axis=None (the default), the result of intp vs. Python integer, changes promotion rules when combined with e.g. float32.

Now, I don't think this will happen a lot, but for count_nonzeros I am not sure I don't think this isn't more user-friendly and an OK and very small inconsistency.
NumPy itself will be fine (because we never only use axis=None), skimage I saw one function which would return float64 rather than a Python float.

(Maybe one of those things where, I feel forcing intp return is great, but I am not sure changing NumPy main namespace for it is worthwhile.)

See numpy/numpy#28615 (comment)