Enhancing the Reasoning Capabilities of Small Language Models via Solution Guidance Fine-Tuning