Towards Fair Federated Learning with Zero-Shot Data Augmentation