Preventing Catastrophic Forgetting: Behavior-Aware Sampling for Safer Language Model Fine-Tuning