llama.cpp b9776 fixes Vulkan overflow with bias-before-softmax patch

The latest release of llama.cpp, tagged b9776, addresses a critical numerical stability issue in the Flash Attention implementation for Vulkan compute shaders. By reordering the bias addition before the softmax operation, the patch prevents overflow errors that could corrupt inference on AMD and other Vulkan-compatible GPUs. This fix is especially relevant for users running large 70B+ parameter models where attention scores are large.

The release was published by the ggml-org maintainers and includes expanded build matrix support: macOS Apple Silicon (both with and without KleidiAI acceleration), iOS XCFramework, Linux across x64/arm64/s390x with CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL FP32/FP16, Windows with CUDA 12/13, OpenCL Adreno, Vulkan, OpenVINO, and HIP, plus Android arm64 CPU and openEuler with Ascend NPU support. This wide coverage makes b9776 one of the most portable versions of llama.cpp to date.