llama.cpp b9980 fixes server multimodal detection and adds Windows CUDA 13 support

The llama.cpp project released version b9980, which includes a fix for the server component regarding multimodal model detection. When the `--no-mmproj-auto` flag is used to explicitly disable multimodal projection, the server no longer declares image or audio inputs as supported modalities on the `/v1/models` endpoint.

• macOS Apple Silicon and Intel builds are available, though KleidiAI support remains disabled.
• Linux binaries cover CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL variants for x64, arm64, and s390x architectures.
• Windows updates include new CUDA 13 support alongside existing CUDA 12, OpenCL Adreno, Vulkan, OpenVINO, SYCL, and HIP options.
• Android arm64 CPU builds and the standalone UI are also provided in this release.