Policy-Guided Causal State Representation for Offline Reinforcement Learning Recommendation