Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks