Optimal Training in large TDD multi-user downlink systems under zero-forcing and regularized zero-forcing precoding