Improving Direct Persian-English Speech-to-Speech Translation with Discrete Units and Synthetic Parallel Data