Improving training of likelihood-based generative models with Gaussian homotopy