Quality versus efficiency in document scoring with learning-to-rank models