xgboost wheel should be built and tested and used separately

[pseudotensor]

To speed-up build process:

1) Wheel should always be build from https://github.com/h2oai/xgboost/tree/h2oai and pushed to private site.

2) We need a separate jenkins job that then takes that wheel and tries to build h2o4gpu and make test on it. Once it's confirmed as a good run, we push the private wheel to public space.

3) We just need then to remove the xgboost build process from the normal build process. We can keep it as present but inactive unless they explicitly call it. This way (in case s3 or whatever is down) or they want to use their own xgboost version. We can keep xgboost as submodule, but just not used.

etc.

[pseudotensor]

@abal5 says @mmalohlava wants to ensure xgboost versioned before we push wheels to s3 for use by build process in h2o4gpu etc.