CoDe-R: Refining Decompiler Output with LLMs via Rationale Guidance and Adaptive Inference