llama.cpp b9949 release adds cluster-parallel decode FA for Adreno

The llama.cpp project has released version b9949, introducing a new OpenCL feature for Adreno GPUs. This update implements cluster-parallel decode Flash Attention (FA) to improve performance on compatible hardware.

• macOS Apple Silicon and Intel binaries are available.
• iOS XCFramework is included in the release.
• Linux builds cover CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL variants for x64, arm64, and s390x architectures.
• Windows support includes CPU, Vulkan, HIP, OpenVINO, SYCL, and CUDA 12/13 binaries.
• Android arm64 CPU builds are provided.
• KleidiAI support on macOS Apple Silicon is disabled in this version.