llama.cpp release b12361 adds new binaries and moves MOE cache code

The ggml‑org team released llama.cpp version b12361. The update moves the MOE‑cache code out of the internal graph handling and provides new pre‑built binaries for macOS (Apple Silicon and Intel), iOS, and several Ubuntu flavors, including CPU and Vulkan builds. A KleidiAI‑enabled macOS Apple Silicon binary is listed but currently disabled. All files are available for download from the GitHub release page.