Instance-Dependent Regret Bounds for Learning Two-Player Zero-Sum Games with Bandit Feedback