Reasoning or Retrieval? A Study of Answer Attribution on Large Reasoning Models