Better Full-Matrix Regret via Parameter-Free Online Learning