Support end-to-end ML workflows

• End to end ML training workflow for the revert risk model
• research_ml module for shared ML abstractions, support for distributed/local training and prediction using xgboost
• ML workflow notebook as an example using research_ml to train models interactively in notebooks

A number of other changes

• introduced a base features step/job, in anticipation of separating the computation of the ML features and generation of a training dataset
• updated the risk observatory batch prediction
• move support for repartition into stratified sample transformation