Certifying the Fairness of KNN in the Presence of Dataset Bias