Improving Mathematical Reasoning Capabilities of Small Language Models via Feedback-Driven Distillation