Smoothing Policy Iteration for Zero-sum Markov Games