Stabilizing Dynamical Systems via Policy Gradient Methods