llama.cpp adds experimental SM120 CUTLASS MoE prefill

llama.cpp is gaining experimental CUDA kernels built on NVIDIA's CUTLASS library for mixture-of-experts prefill on SM120 GPUs, covering the MXFP4 and NVFP4 low-precision formats.

Contributor leonardHONG moved weight repacking so it runs only after a full tensor has reached the GPU, including via chunked uploads. An earlier timing would have broken that path.

JohannesGaessler said gains in the 35 to 40 percent range would make the work far more worthwhile than the roughly 10 percent cited so far. He asked that integration wait on a planned refactor of how quantized matrix-vector kernels are configured, which he called relatively hard to extend modularly for hardware-specific tuning. The kernels remain experimental while that groundwork proceeds.