Pdp v3

[RobertSamoilescu]

Splits the implementation into two classes: PartialDependence and TreePartialDependence.

The PartialDependence accepts a a prediction function similar to ALE, thus can be applied to any black-box model. In the background performs the same computation as sklearn with method='brute'.

The TreePartialDependece is a dedicated for some tree-based sklearn models for faster computation. In the background it performs the same computation as sklearn with method='recursion', kind='average', response_method='decision_function'.

Some advantages of this approach:

• There is no need for the arguments response_method and method anymore which were a bit confusing.
• Similar split as KernelShap and TreeShap -- consistency.
• Simiplifies considerably the sanity checks.