Towards Reliable Empirical Machine Unlearning Evaluation: A Game-Theoretic View