On the Convergence Rates of Policy Gradient Methods