Scaling Laws for Forgetting When Fine-Tuning Large Language Models