test: make strided out tests compare against an independent reference

[antonwolfy]

Four tests that check out= with strided arrays pass the same out to both mkl_fft and numpy.fft:

result = getattr(mkl_fft, func)(x, axis=axis, out=out)
expected = getattr(np.fft, func)(x, axis=axis, out=out)
assert_allclose(result, expected)

Both calls return out, so result is expected and the assertion compares out with itself. The tests pass even if mkl_fft writes nothing, or garbage, into out. test_fft_out_strided also rebinds out to a contiguous array, so despite its name it never tests a strided out.

The affected tests cover 20 parametrized cases, all added in #157:

• test_fft1d.py: test_fft_out_strided, test_rfft_out_strided, test_irfft_out_strided
• test_fftnd.py: test_out_strided

Changes

• Compute the NumPy reference without out, as NumPy's own test_pocketfft.py does.
• Assert that out is returned.
• Make out in test_fft_out_strided an actual strided view.
• Take out as a view of a sentinel-filled buffer and check that nothing outside the view is written to.

Verification

• With master's code, all 20 cases pass, also over 20 repeated runs, and so does the full suite.
• Mutation check: I replaced fft/ifft/rfft/irfft/fftn/ifftn with a stub that returns out untouched. The current tests still pass all 20 cases; the updated tests fail all 20.

This is a test-only change, so there is no CHANGELOG entry. Found while reviewing #387.