Reshard to 10GB chunks (all tensors preserved)

Fix: lm_head + dequant FP8->bf16 + scale slicing + config promotion

Upload folder using huggingface_hub

initial commit