SHAP scores for feature importance

[shuttie]

Now we emit splits as feature importance for all the boosters. But they have some drawbacks:

• they depend on number of trees and tree depth
• features with high cardinality will more frequently be part of split
• hard to compare importance between boosters and different runs

We propose to compute SHAP scores for feature importance:

• a linear estimation on how each feature contributes to the final metric
• same implementation for all the boosters
• does not depend on # of trees and tree depth

TreeSHAP is a major PITA (as shap4j does not support windows, and requires python-specific pickling), so we can just implement KernelSHAP and call it a day:

• there should be a time budget for the whole MC sampling during KernelSHAP evaluation
• we may also emit dispersion as a measure of confidence in the estimation of each feature.