llama.cpp b9993 Adds Support for Hy3 Model Architecture

The latest release of llama.cpp (b9993) introduces support for the Hy3 model architecture, which is based on Tencent's Hunyuan 3. This new addition includes a MoE decoder stack with per-head Q/K RMSNorm and an advanced router mechanism. Local AI operators should be aware that this update enhances compatibility with cutting-edge models like MiniMax-M3 and DiffusionGemma, offering improved performance and accuracy for large-scale language tasks. However, it's crucial to verify if your existing scripts or configurations are compatible with the new Hy3 architecture to avoid potential issues.