* feat(minimax_m3): serve MiniMax-M3-NVFP4 single-GPU (block-sparse GQA + MXFP8 dense + swigluoai NVFP4 experts)
Co-authored-by: Shuo Yang <andy_yang@berkeley.edu>