The Devil is in the Data: Learning Fair Graph Neural Networks via Partial Knowledge Distillation