refactor: np.where(cond) -> np.asarray(cond).nonzero()

[wpbonelli]

The docs for np.where() (https://numpy.org/doc/stable/reference/generated/numpy.where.html) suggest to prefer nonzero() over where() without x and y arguments. In the spirit of defensive programming I included np.asarray(cond) even where cond is already an array.

This PR also fixes a bug I introduced in the model splitter in #2124: while E711 (https://www.flake8rules.com/rules/E711.html) dictates comparisons to None should use identity rather than equality, this rule should not be applied to NumPy array selection conditions as it will change the semantics:

>>> a = np.array([None, None])
>>> a[a != None]
array([], dtype=object)
>>> a[a is not None]
array([[None, None]], dtype=object)

Unrelatedly, mark test_mt3d.py::test_mfnwt_keat_uzf() slow, it should not be included in smoke tests (this was causing the optional dependency CI tests to fail due to timeout). And clean up some unused imports in conftest.py.

[mwtoews]

While np.asarray() isn't always necessary for objects that are already arrays, it shouldn't have any performance impacts (arrays are copied only if necessary, otherwise it just passes the object pointer, thus fast). All other changes look fine to me.