On the Complexity of Policy Iteration