Geometric Variance Reduction in Markov Chains: Application to Value Function and Gradient Estimation