Q-learning with Adjoint Matching