Multi-Bellman operator for convergence of $Q$-learning with linear function approximation