Federated Unlearning with Gradient Descent and Conflict Mitigation