Endowing Pre-trained Graph Models with Provable Fairness