Self-Supervised Reinforcement Learning with dual-reward for knowledge-aware recommendation