Similar to #210, add option to use or not efficient attention to AutoencoderKL
Similar to #210, add option to use or not efficient attention to AutoencoderKL