Linear Convergence of Natural Policy Gradient Methods with Log-Linear Policies