Pitfall of Optimism: Distributional Reinforcement Learning by Randomizing Risk Criterion