llama.cpp b9986 fixes reasoning leak and Nemotron Nano v2 regression

The llama.cpp project released version b9986, addressing specific chat template issues. The update includes a fix for reasoning leaks that occurred with force-opened bare templates.

• Fixes reasoning start tag inference by trimming trailing whitespace to prevent content swallowing.
• Resolves a regression affecting Nemotron Nano v2 models.
• Provides binaries for macOS (Apple Silicon and Intel), iOS, Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Android, Windows (CPU, CUDA 12/13, Vulkan, OpenCL, OpenVINO, SYCL, HIP), and openEuler.

This release ensures correct template handling for specific model architectures and provides updated binaries across supported platforms.