CoEvol: Constructing Better Responses for Instruction Finetuning through Multi-Agent Cooperation