Enhancing Code Generation Performance of Smaller Models by Distilling the Reasoning Ability of LLMs.