Model 4 bit inference 4.2x faster than 16 bit with full HF support

Release notes: https://github.com/TimDettmers/bitsandbytes/releases/tag/0.4...