llama.cpp b9906 fixes ROCm build errors and provides binaries for multiple platforms

The llama.cpp project released version b9906, addressing a critical build failure on ROCm/clang 22 where -ffast-math incorrectly disabled infinity handling. The update adds the -fno-finite-math-only flag to resolve this issue and prevent compilation errors in continuous integration.

• Fixes ROCm build errors by re-enabling INFINITY/NaN support alongside -ffast-math.
• Provides binaries for macOS (Apple Silicon and Intel), iOS, Linux (CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL), Android, Windows (CPU, CUDA 12/13, Vulkan, OpenCL, OpenVINO, SYCL, HIP), and openEuler.
• Includes a standalone UI build for general use.

This release ensures compatibility with newer ROCm toolchains while maintaining broad hardware support across major operating systems and accelerators.