add wrappers for np.fft

[ev-br]

Is mostly straightforward, as pytorch API is nearly identical indeed. The main difference seems to be that numpy.fft returns 64-bit transforms for all input dtypes. We now have set_default_dtype though, so we probably want fft routines to react to it:

>>> a = np.arange(5, dtype=np.float32)
>>> np.fft.fft(a).dtype == np.complex128            # this is the (numpy) way
True
>>> np.set_default_dtype("pytorch")
>>> np.fft.fft(a).dtype == np.complex64

[ev-br]

Had to bump the tolerance. Not much we can do I guess, this falls straight through to pytorch.

[lezcano]

Cool!