Fishing for Magikarp: Automatically Detecting Under-trained Tokens in Large Language Models