Leveraging Unit Language Guidance to Advance Speech Modeling in Textless Speech-to-Speech Translation