Global Convergence of Policy Gradient Methods for Linearized Control Problems