NyoomFloat12: Accelerating LLM Inference via Lossless 12-Bit Weight Compression alphaXiv