[MODEL] add HunyuanOCR quantization support

[ZX-ModelCloud]

Summary

• register native hunyuan_vl checkpoints with a dedicated HunyuanOCR quantization definition
• quantize the language backbone while keeping the vision tower, embeddings, norms, and rotary module in their source dtype
• build multimodal calibration inputs through the processor's official chat template and OCR prompt
• canonicalize lazy checkpoint materialization devices and pass explicit module paths
• preserve nested multimodal RoPE parameters across config normalization and save/reload
• add focused adapter, calibration, config-regression, and ModelTest coverage, and document HunyuanOCR support

Validation

• pytest -q tests/test_hf_config_compat.py tests/test_hunyuan_ocr_support.py tests/models/test_hunyuan_ocr.py::test_prepare_hunyuan_ocr_dataset_reuses_shared_dataset
• focused Ruff, format, compile, and git diff --check validation
• fast and full HunyuanOCR ModelTest quantize/save/reload/evaluation paths
• real-image OCR generation after loading the saved quantized checkpoint with BACKEND.AUTO