SAMG: State-Action-Aware Offline-to-Online Reinforcement Learning with Offline Model Guidance