How Does Value Distribution in Distributional Reinforcement Learning Help Optimization?