support base_margin for xgboost

[bblodfon]

Implement #205

[bblodfon]

https://xgboost.readthedocs.io/en/stable/tutorials/intercept.html#intercept => I think here it says that it can be used also for classif and that it can be also a matrix in case of multi-output/multi-class, I would just add it for binary classification so that the base_margin can be in there - @sebffischer what do you think?

[bblodfon]

@adibender I have tested this, you change the base_margin response predictions do change

[sebffischer]

I don't know how base_margin works, so I just assume this is correct. Just a minor detail about efficiency

[bblodfon]

@sebffischer I added support also for binary classification and tested it. Pretty much base_margin is a baseline/offset prediction where you "boost" from. For regression it's a continuous response, for binary classification it can be any vector again (user is responsible to denote which feature should be used and what values are inside there, eg if the feature has NAs, training won't work)