VAEBM: A Symbiosis between Variational Autoencoders and Energy-based Models