Key-Point-Driven Mathematical Reasoning Distillation of Large Language Model