Build universal wheels using GPU CI

[RAMitchell]

@hcho3 where did we get to with uploading GPU enabled wheels to pypi? I will be releasing a blog post soon it would be nice to have this ready if possible.

[hcho3]

@RAMitchell The current CI already tests the universal wheels on Linux? Yes, I'll upload the wheels on PyPI this weekend. Since Mac and Windows are not supported, we'll have to name the wheels "manylinux". This way, users will fall back to source tarballs for OSes without wheels

[hcho3]

And let us transition to a CI setup in the future where you have multiple slaves, Linux and Windows. (I think this can happen in a month or so.) This way, we can generate Windows wheels. Only complication is that NVidia docker is Linux only, so we'll have to write scripts specific to Windows

[RAMitchell]

Sounds good. I did a bit of testing today, note that the 9.1 wheels throw a runtime error when GPU algorithms are called on systems with cuda 8.0/9.0. We should upload the 8.0 wheel for maximum compatibility, this works fine on all systems I tested.

This PR is basically just to make the CI artifacts directly usable on a python 3 system.

[hcho3]

@RAMitchell Do we really need the --universal flag? It appears to cause test failure (https://xgboost-ci.net/blue/organizations/jenkins/xgboost/detail/PR-3424/1/pipeline). I think we can simply use the artifacts given by the latest successfully Jenkins test, e.g. https://xgboost-ci.net/job/xgboost/job/master/47/.

EDIT. I plan to use the wheels from https://xgboost-ci.net/job/xgboost/job/release_0.72/5/, which has the stable 0.72 version code.

[RAMitchell]

The test failures just due to a file having its name changed - I will fix this. You should be able to use the artifacts that already exist but you will have to rename them to work on python 3. e.g.

xgboost-0.72-py2-none-any.whl -> xgboost-0.72-py2.py3-none-any.whl

Otherwise pip thinks the py2 wheel is incompatible with python 3.

[codecov-io]

Codecov Report

Merging #3424 into master will not change coverage.
The diff coverage is n/a.

@@            Coverage Diff            @@
##             master    #3424   +/-   ##
=========================================
  Coverage     45.55%   45.55%           
  Complexity      188      188           
=========================================
  Files           166      166           
  Lines         13007    13007           
  Branches        442      442           
=========================================
  Hits           5925     5925           
  Misses         6881     6881           
  Partials        201      201

Impacted Files	Coverage Δ	Complexity Δ
python-package/xgboost/core.py	83.45% <0%> (ø)	0% <0%> (ø)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 8bec8d5...c4d1c06. Read the comment docs.