FEA implement a recalibrate parameter for classifier using sampling

[glemaitre]

Implement a new parameter recalibrate to correct for sampling bias introduced due to sampling in ensemble classifiers

TODO:

• implement the same trick for all ensemble classifiers
• add more documentation on how the factor is computed since it is not trivial not correct for each learner instead of averaging
• add an entry in the user guide
• implement an example showing the behaviour via the calibration display
• check a bit more to understand why boosting does have the same issue. Intuitively, since we are boosting randomly on sample, the classifier should be calibrated by increasing the number of weak learner.