Policy Gradient Methods for the Noisy Linear Quadratic Regulator over a Finite Horizon