Assessing the Chemical Intelligence of Large Language Models