llama.cpp b10447 redesigns server threading with yield_to_queue overhaul

ggml-org released llama.cpp b10447 on August 15, introducing a redesigned yield_to_queue thread model for the server component. The change, tracked in PR #27133, replaces the previous scheduling approach with a more efficient cooperative queue mechanism that reduces contention and improves throughput when handling multiple inference requests simultaneously. This is a core infrastructure upgrade for self-hosted LLM serving.

The release also shifts common_speculative_process execution to a new worker-swap design, where speculative decoding runs in worker threads that exchange with the main thread. This separation improves parallelism and reduces stalls during draft-model verification. The update ships with extensive prebuilt binaries covering macOS (Apple Silicon and Intel), iOS, Linux (CPU, Vulkan, ROCm 7.14, OpenVINO, SYCL), Android, Windows (CUDA 12/13, Vulkan, OpenCL, SYCL), and openEuler — ensuring developers can adopt the improvements across edge devices to data-center GPUs.