Step-level Value Preference Optimization for Mathematical Reasoning