Assessment of the Generalization Abilities of Machine-Learning Scoring Functions for Structure-Based Virtual Screening