On the Unreasonable Effectiveness of Last-layer Retraining