Global guarantees for policy gradient methods