Locating Information in Large Language Models via Random Matrix Theory