What we got wrong about scaling laws for fine-tuning