Beyond Prioritized Replay: Sampling States in Model-Based RL via Simulated Priorities