Graph Unlearning with Efficient Partial Retraining