Add hunyuan-dense so Hy-MT2 GGUFs can run

[sdcb]

Summary

• Register hunyuan-dense (Hy-MT2 / Hunyuan dense text) through the architecture + chat-protocol seams. Official Q4 GGUFs previously failed at load with Unsupported architecture.
• Forward matches llama.cpp's hunyuan-vl text graph: RMSNorm → QKV → NeoX RoPE → per-head Q/K RMSNorm → GQA → SwiGLU. QK-norm is after RoPE (the opposite of Qwen 3.5).
• Chat template follows tencent/Hy-MT2-1.8B chat_template.jinja (BOS + add_generation_prompt), not llama.cpp's Hunyuan-4B-Instruct hardcoded template, which dropped BOS and produced off-task replies.
• First cut is SingleDevice: no TP, no fused whole-model graph, no continuous batch. CPU ggml is slower than llama.cpp on this path; correctness was checked on Hy-MT2-1.8B-Q4_K_M (Hello → 你好; official translation prompt produces Chinese).

Test plan

• HunyuanDenseChatTemplateTests (BOS, system, multi-turn, no-tools protocol)
• TensorParallelSupportGateTests / SkillCapabilityConsistencyTests still pass with the new registration
• --backend cpu / ggml_cpu load + translate on Hy-MT2-1.8B-Q4_K_M.gguf
• GPU path (ggml_cuda / cuda) — not exercised on the authoring machine
• Long-context / chunked prefill (model advertises 262k; KV grows on demand)