llama.cpp b9925 adds CUDA f16->f16 GGML_OP_SET_ROWS support

The llama.cpp project has released version b9925, which includes a specific update to its CUDA backend. This release adds support for the f16-to-f16 GGML_OP_SET_ROWS operation, addressing a gap in GPU kernel coverage.

• macOS Apple Silicon and Intel binaries are provided, with KleidiAI explicitly disabled on Apple Silicon.
• Linux builds cover CPU (x64, arm64, s390x) and various accelerators including Vulkan, ROCm 7.2, OpenVINO, and SYCL FP32/FP16.
• Windows releases include CPU, OpenCL Adreno, Vulkan, HIP, OpenVINO, SYCL, and CUDA 12/13 variants.
• Android arm64 (CPU) and iOS XCFramework binaries are available for mobile deployment.
• openEuler support is partially disabled, with specific builds retained for x86 and aarch64 architectures using ACL Graph.

This update ensures broader hardware compatibility for llama.cpp users by enabling the missing set rows operation on CUDA devices and providing precompiled binaries across major operating systems and GPU vendors.