Why long model-based rollouts are no reason for bad Q-value estimates