Support for FP8 quantization

[rwl4]

With the release of the new Mistral NeMo 12B model we now have weights that were pre-trained with FP8. It would be great if Unsloth could support 8bit as well as the existing 4bit training so we could do training without any quantization related loss.