Black-Box Opinion Manipulation Attacks to Retrieval-Augmented Generation of Large Language Models