Shaping Belief States with Generative Environment Models for RL