BUG: scalar integer negative powers gave wrong results.

[ewmoore]

np.int8(4)**np.int8(-4) gave inf. Since we return a float or
double for these cases, just call those power routines. This
prevents integer overflow from resulting in a divide by zero and
the old nonsense result.

[ewmoore]

CI failure is curl failing to download python 3.5. So unrelated to these changes.

[charris]

Hmm, arrays just convert inf to 0, which is the standard cast for integers.

In [3]: array(4, dtype=int8) ** -4
Out[3]: 0

also integer result rather than float. So this is a case where numpy scalars behave differently from array scalars. Apart from that inconsistency this looks good to me, but I wonder if we want to stick with it in the long run. Perhaps a note to the mailing list would elicit some discussion.

[charris]

OTOH, it makes it consistent with Python

In [1]: 2 ** -2
Out[1]: 0.25

So let's put this in. The inconsistency is still a bit bothersome.

[charris]

Thanks @ewmoore .