Test all Numba decorated functions with care

[hakonanes]

The lesson from #494 is that all Numba decorated functions must be tested with care, meaning not just via py_func() or physical examples, but "indirectly". I'm not sure how to do this systematically, but I hope this issue can guide us in that direction.

The issue refers to different results on Windows and Linux produced by a Numba decorated function for converting a Euler angle triplet to a quaternion. It was fixed by replacing array creation in two steps

kikuchipy/kikuchipy/_rotation/__init__.py

Lines 65 to 69 in 5311982

	rotation = np.zeros(4)
	rotation[0] = c * np.cos(sigma)
	rotation[1] = -s * np.cos(delta)
	rotation[2] = -s * np.sin(delta)
	rotation[3] = -c * np.sin(sigma)

with creation in one step

kikuchipy/kikuchipy/_rotation/__init__.py

Lines 65 to 68 in adfc6f0

	rotation = np.array(
	(c * np.cos(sigma), -s * np.cos(delta), -s * np.sin(delta), -c * np.sin(sigma)),
	dtype=np.float64,
	)

For some reason, the two-step creation gave garbage results on Windows but not on Linux, while the latter gave same results on both OS. This was not caught in testing, but is now, by testing another function which calls this function.

[hakonanes]

Tips for testing Numba functions have been added to the contributing guide in #499.