The Super Weight in Large Language Models