llama.cpp sm120 CUDA build hits Windows snag

This looks less like llama.cpp being fundamentally broken and more like Blackwell/CUDA support still settling on Windows. NVIDIA's CUDA 12.8 docs add SM_120 compiler support, so the architecture itself is real; the rough edge is the surrounding build stack and kernels. llama.cpp's build docs already cover CUDA, non-native builds, and explicit CMAKE_CUDA_ARCHITECTURES, which gives supported escape hatches when auto-detection misbehaves. Other Windows reports on RTX 5090-class hardware show CUDA builds compiling and detecting compute capability 12.0, so this feels like a fragile compatibility pocket rather than a total lack of support. Vulkan is the pragmatic fallback if you want stable local inference now instead of spending time on the newest CUDA edge cases.