Implement take and take_along_axis for time dtypes

[Christian Bourjau (cbourjau)]

No description provided.

[SamuelLessmannQC]

Is there a specific reason not to reshape like for repeat as done here?

This fails:

@pytest.mark.parametrize("dtype", ["datetime64[s]", "timedelta64[s]"])
def test_take_axis_none_flattens(dtype):
    np_arr = np.asarray(
        [
            [1, "NaT", 0],
            [10, 20, 30],
        ],
        dtype=dtype,
    )

    def do(npx):
        indices = np.asarray([4, 1, -1], dtype=np.int64)
        return npx.take(npx.asarray(np_arr), npx.asarray(indices))

    np.testing.assert_array_equal(do(ndx).unwrap_numpy(), do(np))

[Christian Bourjau (cbourjau)]

take[_along_axis] and repeat have quite different semantics with regard to their axis argument in the array-api. Some of the test cases, as well as your example, are not specified by the array-api. This was a good catch, though! We didn't validate the arguments properly and had diverging behavior compared to NumPy. I added proper validation and updated the test cases.