llama.cpp b9938 enables fast math for AMD/HIP via -funsafe-math-optimizations

The llama.cpp b9938 release updates the ggml-hip backend to enable the -funsafe-math-optimizations compiler flag, allowing AMD/HIP builds to use fast math optimizations previously exclusive to CUDA.

This change provides a speedup for AMD hardware without causing NaN issues or compilation failures associated with the stricter -ffast-math flag. The release includes binaries for macOS (Apple Silicon and Intel), Linux (CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL), Android, Windows (CPU, OpenCL Adreno, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), and openEuler.

The update ensures parity in optimization capabilities between NVIDIA and AMD backends while maintaining numerical stability for masking operations.