External Reasoning: Towards Multi-Large-Language-Models Interchangeable Assistance with Human Feedback