Accelerating Primal-dual Methods for Regularized Markov Decision Processes