Step-KTO: Optimizing Mathematical Reasoning through Stepwise Binary Feedback