Towards Last-layer Retraining for Group Robustness with Fewer Annotations