Robust Counterfactual Explanations on Graph Neural Networks