Currently, sorting and comparison of subarrays is wrong in many cases.

This was discovered because trying to sort arrays with subarrays leads to different results on little-endian systems vs big-endian:

>>> z = np.array([([3,3],), ([-1,0],)], dtype=[('v', '>i4', (2,))])
>>> z[0] < z[1]

On little endian, the last line returns True. On big-endian, it returns False.

Here's what's happening: In VOID_compare, this line, see that non-structured voids are compared by doing cmp(val1.tobytes(), val2.tobytes(). In other words, it looks like we treat subarrays as if they were plain 'np.void' types and use a simple string compare. That's why endianness mattered, but it also means that a lot of other cases are wrong since we are comparing the byte values rather than the numerical values.

I think we need to add an extra block there to take care of subarrays.