Fix building wheels on OSX

[menshikh-iv]

piskvorky/gensim#2794

[piskvorky]

@menshikh-iv thanks – is there a TL;DR version of what was wrong?

[menshikh-iv]

@piskvorky

Hidden fails 1

The story starts from 1009a3f, 1 year ago this was needed for release wheels (CI doesn't give them if tests failed) and this was reverted only several days ago (when I dug into CI logs and found this).

For this reason, some of the wheels can be broken, because we don't track test fails.

How this works now:

• test fail -> CI job fail
• wheels produced in any case (no matter, is tests failed or not), to avoid || true situation in the future.

Hidden fails 2

According to the changes above, I created piskvorky/gensim#2791 to fix some of them & checking build. So, this partially helps (at least with our tests), but doesn't fix all. Others wired problem (especially with OSX) still persist here.

Travis updates

From start, wheels for osx was broken, I'm using piskvorky/gensim#2794 as indicator + carefully read logs (we have a batch of wired exceptions here)

• Travis update CI file directives, so, I updated them according to modern guidelines
• OSX VMs changed, this was the main reason for "broken osx wheels". I investigated, how this works in numpy and scikit-learn wheels and apply these approaches to the current repo.

[piskvorky]

Thanks.

From start, wheels for osx was broken

How do you mean? Many OSX wheels worked, when I tested here.

[menshikh-iv]

How do you mean?

When I start working on this PR, OSX can build only 3.5 wheels (with batch of failed tests after, but it doesn't matter), 3.6 and 3.7 doesn't produce a wheel at all (broken on compilation stage with wired errors)