Token-Supervised Value Models for Enhancing Mathematical Reasoning Capabilities of Large Language Models