Mechanistic Analysis of Catastrophic Forgetting in Large Language Models During Continual Fine-tuning