Trustworthy Alignment of Retrieval-Augmented Large Language Models Via Reinforcement Learning