Proactive Privacy Amnesia for Large Language Models: Safeguarding PII with Negligible Impact on Model Utility