Bringing together invertible UNets with invertible attention modules for memory-efficient diffusion models