On the convergence of policy iteration for controlled diffusions