Enhancing Vision-Language Models with Chain of Manipulations: A Leap Towards Faithful Visual Reasoning and Error Traceability