Refactor stacking estimators with class inheritance

[kota7]

Currently we have four stacking-based estimators: StackingRegressor and StackingClassifier, and CV variant of each. They are independently coded, while core steps are common.

This makes enhancement work time-consuming, since contributors often need to make similar update on four source code. I would suggest refactoring the four estimators with appropriate class inheritance structure. This will make issues like #439 and #255 much simpler to implement.

[rasbt]

Totally agree. They were implemented one at a time, over time, and the API changed quite a bit based on what's useful and what not. Now after 1-2 years that it's kind of stabilized and they remain pretty similar "core"-wise, it'd be good to consolidate them via a parent Stacking class