llama.cpp b9881 release enables -ffast-math for HIP builds

The llama.cpp project has published the b9881 release, providing binaries for macOS, Linux, Windows, Android, and openEuler across CPU, GPU, and specialized hardware backends.

• ggml-hip: enable -ffast-math for HIP builds
• macOS Apple Silicon (arm64) KleidiAI support is disabled
• Ubuntu x64 builds now include OpenVINO 2026.2.1 and SYCL FP32/FP16 variants
• Windows binaries add CUDA 13 support alongside existing CUDA 12, Vulkan, and ROCm options
• openEuler x86 (310p) and aarch64 builds are available, while the previous openEuler pull request is marked disabled

This release updates the available precompiled binaries for developers targeting diverse hardware architectures.