CI: NumPy is being upgraded

[TomAugspurger]

Looks like NumPy is being upgraded behind our back on master: https://travis-ci.org/pandas-dev/pandas/jobs/322904433#L1313

That's what's causing the failures on the 0.22 backports branch. Master is currently failing with NumPy 1.10, but we aren't catching it.

I'll look into it now.

cc @jreback

[TomAugspurger]

I may just use this issue for fixing the test failures in https://travis-ci.org/pandas-dev/pandas/jobs/322911128#L1783

At the moment this seem to be related issues seem to be https://github.com/numpy/numpy /pull/8231 and #15363 (power of integer array to integer array). This test is skipped on NumPy>=1.12

(Pdb) df.A.values[:5]
array([17, 92, 17, 79, 54])
(Pdb) other.A.values[:5]
array([17, 92, 17, 79, 54])
(Pdb) df.A.values[:5] ** other.A.values[:5]
array([-9223372036854775808, -9223372036854775808, -9223372036854775808,
       -9223372036854775808, -9223372036854775808])

So I'm tempted to just remove pow from that test.

[jreback]

oh, yes we DO skip on pow in test_binary_ops already (though its hacky). prob same issue. my fix is too extreme.

[TomAugspurger]

Yeah, it's just pow. Opened #18995 so that it's easy to backport.