Persistently Trained, Diffusion-assisted Energy-based Models