Scaling Relationship on Learning Mathematical Reasoning with Large Language Models