Measuring Social Norms of Large Language Models