NVFP4 models land on native Windows

NVFP4 is the "killer app" for the RTX 50-series, offering a rare win-win of massive VRAM savings without the typical accuracy degradation of 4-bit integer formats. Native Windows support removes the significant "WSL tax" for developers, allowing direct GPU access without the complexity of virtualized environments. Building with CUDA 12.8 is critical, as newer versions currently break Blackwell-specific MMQ kernels in llama.cpp. This structural shift to FP4 leverages Blackwell hardware to maintain near-FP8 accuracy, enabling 70B+ parameter models to run on consumer-grade 16GB VRAM cards.