Rewrite assert np.* tests to use numpy.testing

[timhoffm]

PR Summary

Use np.testing.assert_allclose() instead of assert np.allclose() and similar because the numpy assertions will give context on what fails whereas pytest cannot introspect np.allclose() and will just notify an overall failure.

[anntzer]

Perhaps use assert foo == pytest.approx(bar) (https://docs.pytest.org/en/latest/reference.html#pytest-approx) instead of assert_allclose? This way you get the nice pytest-like syntax, context info on failure, and it's also clearer (to me?) what value is the "reference" (result = approx(reference)).

(Also there's a bunch of "allclose-asserters" (https://docs.scipy.org/doc/numpy/reference/routines.testing.html#asserts: assert_almost_equal, assert_approx_equal, assert_array_almost_equal, assert_allclose, assert_array_almost_equal_nulp, assert_array_max_ulp), I'm sure there may be cases where one clearly wants one of them, but in matplotlib I'd guess most often that was not really thought out, so pytest.approx also just provides a nice, standard way to spell that.)

[timhoffm]

pytest.approx gives reasonable error messages for scalars:

>       assert a == pytest.approx(6)
E       assert 5 == 6 ± 6.0e-06
E        +  where 6 ± 6.0e-06 = <function approx at 0x7f4b00930d08>(6)
E        +    where <function approx at 0x7f4b00930d08> = pytest.approx

but for numpy arrays, they are not quite readable (in particular it does not give the position of the error):

        a = np.array([1, 2, 3])
        b = np.array([1, 2, 4])
>       assert a == pytest.approx(b)
E       assert array([1, 2, 3]) == approx([1 ± 1.0e-06, 2 ± 2.0e-06, 4 ± 4.0e-06])
E        +  where approx([1 ± 1.0e-06, 2 ± 2.0e-06, 4 ± 4.0e-06]) = <function approx at 0x7f12827fcd08>(array([1, 2, 4]))
E        +    where <function approx at 0x7f12827fcd08> = pytest.approx

Therfore, I would just stick with numpy.testing for these cases for now. The changes here are an incremental improvement. We have more numpy.testing calls. They could all be changed together later when there's better introspection into numpy arrays from within pytest.

Semi-related: I'm investigating if there are better ways for exact comparison: pytest-dev/pytest#5347

[NelleV]

Thanks @timhoffm !