Manifold Embeddings for Model-Based Reinforcement Learning under Partial Observability