Convergence and Sample Complexity of Policy Gradient Methods for Stabilizing Linear Systems