llama.cpp b9846 patches Vulkan matmul for Asahi Linux users

llama.cpp, the popular C++ inference engine for LLMs, released version b9846 with a targeted fix for the Vulkan backend. The update addresses a matrix multiplication (matmul) regression affecting Asahi Linux on Apple Silicon hardware. Specifically, the release rolls back a BK-loop unroll change that caused performance issues, and instead edits SPIR-V shader code directly to optimize the matmul loop. This ensures that users running local LLMs via Vulkan on Asahi Linux experience smoother, faster inference.

The release also includes minor cleanups like fixing an inline comment and removing trailing whitespace. The b9846 tag is built for multiple platforms: macOS (Apple Silicon and Intel, with optional KleidiAI), Linux (x64, arm64, s390x with Vulkan, ROCm, OpenVINO, SYCL variants), Windows (x64, arm64 with CUDA, Vulkan, OpenVINO, SYCL, HIP), and Android (arm64). The fix is particularly important for developers using AMD GPUs or Apple Silicon with the Vulkan API on Linux, as it restores expected matmul performance without sacrificing cross-platform compatibility.