Leveraging Domain-Unlabeled Data in Offline Reinforcement Learning across Two Domains