Characterizing Attribution and Fluency Tradeoffs for Retrieval-Augmented Large Language Models