llama.cpp OpenVINO: Qwen3.5 Support & Major Memory Optimizations

The llama.cpp OpenVINO backend now supports Qwen3.5 models and introduces significant memory optimizations for GPU users, including a new `GGML_OPENVINO_RELEASE_WEIGHTS` mode that drastically reduces host RSS. These updates also bring improved accuracy and stability across various operations and architectures.