Generate all recommendations in one matrix multiplication

[askeluv]

In our offline evaluation (precision, recall), we currently iterate through all users and run get_recommendations for each user. This is slow.
Instead we should generate the full user x item recommendation set in one matrix multiplication operation. This involves changing get_precision_recall in https://github.com/blockchain-etl/token-recommender/blob/master/wals_ml_engine/trainer/wals.py