llama.cpp b9945 fixes chat-template thinking probe crash

The llama.cpp project released version b9945, which includes a critical fix for the server component's chat-template thinking probe. The change moves the probe inside the init try/catch block to prevent uncaught exceptions.

Previously, models with chat templates that parsed at initialization but failed during parser generation (e.g., using {% call %}) caused std::invalid_argument errors outside the guarded scope. This resulted in llama-cli aborting via SIGABRT instead of gracefully failing to load. The fix ensures apply-time errors are handled consistently with init parse errors.

The release provides binaries for macOS, Linux, Windows, Android, and openEuler across various architectures including CPU, CUDA, Vulkan, ROCm, OpenVINO, SYCL, and OpenCL.