When to Trust Your Simulator: Dynamics-Aware Hybrid Offline-and-Online Reinforcement Learning