Tackling the Generative Learning Trilemma with Denoising Diffusion GANs