Convergence of policy gradient methods for finite-horizon stochastic linear-quadratic control problems