Closing the Gap between TD Learning and Supervised Learning with $Q$-Conditioned Maximization