Self-Assessment Reinforcement Learning for Search-State Awareness in Retrieval-Augmented Large Language Models