Finite-sample Guarantees for Nash Q-learning with Linear Function Approximation