Fairness Testing of Large Language Models in Role-Playing