Variance Control for Distributional Reinforcement Learning