Constrained Reasoning Chains for Enhancing Theory-of-Mind in Large Language Models