llama.cpp b9946 adds Hexagon unary op optimizations and new binaries

The llama.cpp project has released version b9946, introducing significant optimizations for Qualcomm Hexagon processors alongside new build artifacts for multiple platforms.

• Hexagon unary operations now tile wide rows to avoid VTCM overflow and replace divisions with fastdiv.
• The update includes tracing instrumentation and streamlined HVX kernels for improved code generation.
• macOS Apple Silicon binaries are available, though KleidiAI support remains disabled.
• New Linux builds cover CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL targets.
• Windows releases now include CUDA 13 support alongside existing CUDA 12, OpenCL, and HIP options.

This release provides updated binaries for desktop and mobile environments while enhancing performance on Hexagon-based hardware.