How to Fine-Tune Vision Models with SGD