More Benefits of Being Distributional: Second-Order Bounds for Reinforcement Learning