[quantization] Add model_args to PTQConfig for passing model-specific execution metadata

[mhs4670go]

Motivation

We are adding support for multimodal models such as Qwen-VL, which require additional model-specific metadata during quantization and export.

For example, Qwen-VL needs vision_grid_thw to construct position embeddings statically. This information is not part of quantization policy, but rather execution-time metadata required by certain wrappers.

Currently, PTQConfig only supports:

• global defaults (dtype, observer, qscheme)
• per-scope overrides (observer configuration)

There is no clean way to pass such model-specific metadata through the existing
configuration system.

Problem

We need a mechanism to provide additional inputs (e.g., vision grid shape)
to wrappers without:

• polluting HuggingFace model config
• modifying wrapper constructor signatures
• abusing overrides for non-quantization purposes

Proposal

Introduce a new field in PTQConfig:

model_args: Mapping[str, Any] = field(default_factory=dict)

This field is intended to store model-specific execution metadata required by wrappers.

Example Usage

cfg = PTQConfig(
    default_dtype=DType.int(16),
    default_qscheme=QScheme.PER_TENSOR_SYMM,
    wrapper_variant="prefill",
    model_args={
        "vision": {
            "grid_thw": (1, 36, 36),
        }
    },
)

In wrapper:

vision_args = qcfg.get_model_arg("vision", {}) if qcfg else {}
vision_grid_thw = vision_args.get("grid_thw")