Answering Unseen Questions With Smaller Language Models Using Rationale Generation and Dense Retrieval