The Benefits of Model-Based Generalization in Reinforcement Learning