[FEATURE] enable output_margin, pred_contribs in sklearn API

We maintain a slightly adjusted version of xgboost for our own purposes, but one thing that is common are some trivial features that are not exposed. I've mentioned this to @trivialfis before, but wanted to bring up since it's been several months and we are applying latest changes to ours.

Here's what I'd recommend:

1) sklearn.py pred_leaf, pred_contribs, approximate_contribs

diffs here:
(some here, but not all): https://github.com/h2oai/xgboost/commit/bfff9ab50084eaf3311379d7ae194cca3b6f94f0
i.e. at: https://github.com/h2oai/xgboost/blob/bfff9ab50084eaf3311379d7ae194cca3b6f94f0/python-package/xgboost/sklearn.py
i.e. API related ones (not label encoder ones): 

2) output_margin de-hard-code and use:

https://github.com/h2oai/xgboost/commit/355e415c1c8a5a148b5a4fef545d0ccc7b44b29c

As-is, output_margin now can no longer be chosen independently. And pred_contribs etc. are not accessible still.

Start in python-package/xgboost/sklearn.py and compare the referenced h2oai/xgboost commits, focusing on the sklearn API changes rather than label-encoder changes. Trace how output_margin is currently constrained and how pred_leaf, pred_contribs, and approximate_contribs are exposed. Done means these options are independently available through the sklearn API.

Written by the indexing model from the issue text.