llama.cpp b9529 fixes GPU layer bug, expands platform builds

ggml-org has released llama.cpp version b9529, a maintenance update that squarely addresses a bug in the `llama_model::n_gpu_layers()` function (issue #24188). This function is critical for determining how many transformer layers are offloaded to GPU(s), and the bug could lead to improper allocation—especially in multi-GPU setups or when using partial offloading. The fix restores correct layer counting, ensuring that models using GPU acceleration (CUDA, ROCm, Vulkan, etc.) perform predictable and optimized inference.

The release also showcases llama.cpp's commitment to broad platform support. It ships with pre-built binaries for macOS (Apple Silicon and Intel, including a KleidiAI-enabled variant), Linux (CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32), Windows (CPU, CUDA 12 & 13, Vulkan, HIP), Android arm64, and openEuler (x86 and aarch64 with ACL Graph). Additionally, the release includes UI assets and is GPG-signed for verification. This update is particularly relevant for developers deploying LLMs on heterogeneous hardware who rely on precise GPU layer control.