llama.cpp b9966 release with tensor-split regex optimization

The llama.cpp project has released version b9966, featuring a performance optimization for the tensor-split regex patterns. The `llama_meta_device_get_split_state()` function previously recompiled 29 `std::regex` objects on every call, which dominated the decode thread in profiling during -sm tensor mode.

• Marking these patterns as static const ensures they are compiled only once, improving efficiency without changing behavior.
• macOS Apple Silicon (arm64) builds are available, but KleidiAI support is disabled.
• Linux binaries are provided for Ubuntu x64 and arm64 (CPU), s390x (CPU), Vulkan, ROCm 7.2, OpenVINO, and SYCL FP32/FP16.
• Android arm64 (CPU) builds are included.
• Windows releases cover CPU, OpenCL Adreno, CUDA 12/13, Vulkan, OpenVINO, SYCL, and HIP.
• openEuler support is partially disabled, with x86 and aarch64 builds available for specific hardware.

This update provides optimized binaries across multiple platforms and accelerators, addressing performance bottlenecks in tensor splitting.