Add Hunyuan dense V1 (hunyuan_v1_dense): Hunyuan-MT-7B and Hy-MT2-7B

[beshkenadze]

What

Adds Tencent's Hunyuan dense V1 architecture (hunyuan_v1_dense / HunYuanDenseV1ForCausalLM), used by the translation models Hunyuan-MT-7B and Hy-MT2-7B. Ported from mlx-lm's hunyuan_v1_dense.py.

Architecture

A Llama-family dense transformer, closest to the existing Qwen3 path:

• Per-head QK RMSNorm, GQA, SwiGLU MLP, pre/post RMSNorm blocks.
• Tied embeddings (sanitize drops the tied lm_head).
• DynamicNTKAlphaRoPE — the one new utility: rescales the RoPE base once by alpha^(dim/(dim-2)) and reuses the existing freqs-based fast-RoPE path (no sequence-length-dependent resizing). Lives alongside YarnRoPE/Llama3RoPE/ProportionalRoPE in RoPEUtils.swift.

Changes

• Libraries/MLXLMCommon/RoPEUtils.swift: DynamicNTKAlphaRoPE.
• Libraries/MLXLLM/Models/Hunyuan.swift: HunyuanModel / HunyuanConfiguration (flat config decode; accepts head_dim or attention_head_dim; conditional qk-norm).
• Libraries/MLXLLM/LLMModelFactory.swift: register hunyuan_v1_dense; presets hunyuan_mt_7b_4bit/8bit and hy_mt2_7b_4bit/8bit.
• Tests/MLXLMTests/HunyuanTests.swift: config decode, attention_head_dim alias, sanitize, tiny forward, dynamic-RoPE alpha, presets.

Validation

• 6/6 unit tests pass; package builds.

• End-to-end byte-identical parity vs the mlx-lm reference (greedy) on locally-converted 4-bit weights for both models:

  Prompt: Translate the following segment into Chinese, without additional explanation.\n\nIt's on the house.

  model	mlx-lm	Swift HunyuanModel
  Hunyuan-MT-7B-4bit	这顿饭由我们公司承担费用。	这顿饭由我们公司承担费用。
  Hy-MT2-7B-4bit	这顿算店的。	这顿算店的。

  Identical — validates config decode, weight loading, qk-norm, GQA, dynamic RoPE, tied embeddings, and the chat template.

MLX weights: mlx-community/Hunyuan-MT-7B-{4bit,8bit} and mlx-community/Hy-MT2-7B-{4bit,8bit}.