Federated Unlearning: a Perspective of Stability and Fairness