Add K-fold training strategy for XGBoost Ranker

[tomaz-suller]

During experiments, we noticed the strategy suggested in the XGBoost ranker notebook -- of separating a holdout evaluation set to use as training set for the ranker model -- did not provide us any reasonable results. To tackle this and allow us to train the ranker using the complete training set, we developed the scripts in this PR:

• run_train_kfold.py trains recommender models using K-fold cross-validation, and saves the model trained on each fold for later inference;
• run_ranker_dataset.py uses the saved models to compute item features, as well as several other features derived from user interactions and item features.