llama.cpp b9932 disables FA mask_opt on GCN to improve Vulkan performance

The llama.cpp project released build b9932, which includes a Vulkan optimization that disables the FA mask_opt feature on GCN architectures to enhance performance.

• The update re-enables mask opt for attention head sizes greater than 256.
• macOS Apple Silicon builds with KleidiAI are disabled in this release.
• Binaries are provided for Ubuntu (CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL), Windows (CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), Android, and macOS Intel.
• openEuler support is partially disabled, with x86 and aarch64 builds available for specific hardware configurations.