Control Regularization for Reduced Variance Reinforcement Learning