Test NumPy 1.12 and latest in both Travis and tox

[nils-werner]

I have now run into issues with NumPy 1.12 twice: #71 #78. Maybe it makes sense to test multiple Python versions and multiple NumPy versions?

This PR creates a testmatrix in both Travis and tox that tests all combinations of

• Python 2.7
• Python 3.6

and

• NumPy 1.12.0
• NumPy latest

In tox you can run a specific test on a version combination using

tox -e py36-numpy12 -- -k testname

for testname on Python 3.6 and NumPy 1.12.0

[hameerabbasi]

Since we're making extensive use of the __array_ufunc__ interface which was introduced in Numpy 1.13, we should require Numpy 1.13.0 rather than supporting legacy versions. It should be okay, since most packages (at least in the Anaconda ecosystem) have been updated to support Numpy 1.13.

[nils-werner]

Wait but in #71 it was the other way round, no? :-)

[hameerabbasi]

Merging #72 might help fix this... I'll re-run once that's merged. Otherwise I'm just in favor of dropping Numpy 1.12 support.

#71 was different because it was a minor behavioral change leading to a test failure. Here, it's a whole paradigm shift between how Numpy 1.12 and 1.13 handle operations in general.

[hameerabbasi]

We're no longer supporting Numpy 1.12 (see #84 for all the benefits).