Adding python 3.6 builds to travis - #5642

@bsipocz - I wonder if it might make more sense to update the default from 3.5 to 3.6 and add an explicit 3.5 build?

@astrofrog - Sounds like a good plan. We should anyway wait with this PR until conda provides python 3.6 and packages. A related question: should we change the version in appveyor, too?

@bsipocz - I would lean on the side of updating AppVeyor too

Python 3.6 is available on conda defaults now, along with the core packages (numpy, scipy, a few others). It will take them a couple weeks to build out the rest of the packages in defaults: https://groups.google.com/a/continuum.io/forum/m/?utm_medium=email&utm_source=footer#!msg/conda/QJenUjUWKYc/j2DdI9f6EAAJ

Thanks @mwcraig for the heads up. I've just force pushed this, so hopefully travis will work.

I've built 1.0.11, 1.2.2 and 1.3 on python 3.6 for astropy-ci-extras. Will move some of that to the astropy channel on Monday.

Still no h5py on python 3.6...

Nor Matplotlib. I think we'll need to be patient.

This PR also removes the sphinx version limitation (it was needed with <1.3 helpers, see #5568)

I've temporarily added conda-forge as a channel for the py3.6 builds to pick up the latest setuptools. Indeed it helped, and the weird error disappeared, but now there are other mpl related failures.
cc @astrofrog @MSeifert04

https://travis-ci.org/astropy/astropy/jobs/194574394

@bsipocz - can you rebase and try again without conda-forge? We really don't want to be using conda-forge for the CI, mixing conda-forge and defaults leads to all kinds of fun.

@astrofrog - Sure. I suspect it will still fail due to the old setuptools on defaults, and also due to the mpl issues.

It actually seems to have worked, with one probably correct error for a \s escape sequence in a string (should this be a byte string at all? Maybe just a raw string?):

File "/home/travis/build/astropy/astropy/astropy/config/configuration.py", line 601

    content = re.sub(b'\njquery_url\s*=\s*[^\n]+', b'', content)

This PR now includes a temporary xfail for the windows version of the tests that are failing with np 1.12. The xfail can be removed once #5797 is fixed.

🎉 this is all green now, we can start testing with python3.6!

Fantastic! Looked over it once more and think all is OK, so merging!

@mhvk - thanks for merging this! However, I noticed the following issue with this pull request:

• The milestone has not been set

Would it be possible to fix this? Thanks!

If you believe the above to be incorrect (which I - @astrobot - very much doubt) you can ping @astrofrog