Federated Unlearning in the Wild: Rethinking Fairness and Data Discrepancy