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