DFloat11: Fast and memory-efficient GPU inference for losslessly compressed LLMs and diffusion models