Fix integer overflow in np.product

[cgohlke]

On win-amd64-py3.x, test_convolve_fft.test_big_fail fails to raise a ValueError due to an integer overflow in np.product.
The following code raises a RuntimeWarning and allocates huge amounts of RAM:

import numpy as np
from astropy. convolution import convolve_fft
arr = np.empty([512, 512, 512], dtype=np.complex)
convolve_fft(arr, arr)

X:\Python34\lib\site-packages\astropy\convolution\convolve.py:391: RuntimeWarning: overflow encountered in long_scalars
  array_size_B = np.product(arrayshape)*np.dtype(complex_dtype).itemsize

[embray]

Is this in dev or a particular version of Astropy? I didn't find this in my Windows testing.

[cgohlke]

This is v0.4rc1 on win-amd64-py3.x.

[embray]

@astrofrog -- like I said, all the tests passed for me on Windows but that doesn't mean there are no issues

[astrofrog]

@embray - shall we merge this anyway since it can't really cause any harm? Looks like this was submitted from the Github interface which means it's not easy to rebase, so we should add the changelog entry when merging.

[embray]

Sounds fine.