Support multimodal calibration when quantizing only the LLM backbone

Problem

On the latest upstream main (bb291453bc10), it is not possible to calibrate a VLM with real image-text samples when only the LLM backbone is quantized and the vision tower stays in FP16.

Multimodal calibration is only used when --visual_quantization is enabled:

elif visual_quantization is not None:
    image_ds = resolve_dataset(image_dataset, "image")
    batches = _multimodal_calib_dataloader(...)
    mtq.quantize(
        model,
        quant_cfg,
        forward_loop=lambda m: _calibrate_multimodal(m, batches),
    )
else:
    # text-only calibration

Source: quantize.py#L950-L975

When --visual_quantization is not set, passing --image_dataset has no effect and calibration falls back to text-only inputs. This means the LLM quantizers never see the visual-token and multimodal chat-template inputs used in production.

Expected behavior

Please add an independent option, for example:

tensorrt-edgellm-quantize llm \
  --quantization fp8 \
  --image_dataset my_vlm_data \
  --multimodal_llm_calibration

It should:

• Keep the vision tower in FP16;
• Run calibration with image-text inputs through the full VLM;
• Collect activation statistics only for the quantized LLM-side modules;
• Leave existing --visual_quantization behavior unchanged.

This is a Python/ModelOpt calibration-flow issue, not something that can be addressed through a TensorRT plugin.

Start in tensorrt_edgellm/quantization/quantize.py around lines 950-975 and trace how image_dataset, visual_quantization, and the text-only calibration branch are selected. Verify the new independent option calibrates the full VLM with image-text inputs while collecting statistics only for quantized LLM modules, keeps the vision tower in FP16, and leaves visual_quantization behavior unchanged.