Probing the Robustness of Theory of Mind in Large Language Models