Improved sklearn compatibility

[daviddiazvico]

Related issue: #2720

[msftclas]

All CLA requirements met.

[StrikerRUS]

@daviddiazvico First of all, thank you very much for your efforts!

We are aware of the issue you've mentioned, please refer to #2628. However, scikit-learn official development guide doesn't say that it is prohibited to set any params with leading _ in __init__, only the ones with trailing _ are prohibited.

There is an upstream issue in which scikit-learn team are discussing right now how inconsistency between public API tests and development documentation should be fixed: scikit-learn/scikit-learn#16241. Actually, an expected from third-party estimators behavior is described in this comment.

Our sklearn wrapper is quite well tested with time and we are not aware of any confirmed issues with it. So, I don't think we want to refactor it right now. Instead, we'll wait for a fix of scikit-learn/scikit-learn#16241, which I suppose will be in removing confusing public API check.

In addition, I've already prepared for any other decision on scikit-learn/scikit-learn#16241 and prepared a branch a while ago which removes private attributes from __init__ and looks a little bit more complete comparing to your PR: https://github.com/microsoft/LightGBM/compare/sklearn. So, I'm going to close this PR. Thanks again!