Factor out calibration from CalibratedClassifierRatio

[glouppe]

• [x] Factor out calibration from CalibratedClassifierRatio and simplify code into a ClassifierRatio
• [x] Extend sklearn.calibration.CalibratedClassifierCV with distribution-based calibration routines

[glouppe]

Done as part of 5e6a24f0628173e7e3ba20928fce905256fa34ed.

The API is now simpler and fully decoupled:

• carl.learning.CalibratedClassifierCV takes care of calibrating the probability output of a classifier (following the same interface as sklearn.calibration.CalibratedClassifierCV, but with extended support for histogram / kde based calibration).
• carl.ratios.ClassifierRatio uses the probability output to compute the ratio, agnostically of the fact that the probabilities are calibrated or not .
• ClassifierRatio(CalibratedClassifierCV(my_estimator)) can be used to form approximated ratios based on calibrated classifiers. The notebook examples have been updated accordingly.