From Static to Dynamic: Enhancing Offline-to-Online Reinforcement Learning via Energy-Guided Diffusion Stratification