Reward-Free Curricula for Training Robust World Models