Game-theoretic Counterfactual Explanation for Graph Neural Networks