On the Global Convergence Rates of Softmax Policy Gradient Methods