fix: revert to quantized MTP weights (BF16 MTP produces 0% acceptance on quantized backbones)

fix: replace quantized MTP weights with BF16 for higher acceptance rate

Upload folder using huggingface_hub

initial commit