Model-Invariant State Abstractions for Model-Based Reinforcement Learning