Automatically Detecting Under-Trained Tokens in Large Language Models