Learning Energy-Based Models by Diffusion Recovery Likelihood