Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models