PID Control-Based Self-Healing to Improve the Robustness of Large Language Models