Constrained Variational Policy Optimization for Safe Reinforcement Learning