[Fix] Allow InternVL to skip dynamic preprocessing

[zcxGGmu]

Motivation

Issue #3778 asks how to avoid InternVL's internal dynamic resize/splitting when the application needs a stable mapping back to a prepared image, e.g. grounding coordinates.

Refs #3778

Modification

• Add skip_preprocess=True support for the legacy InternVL dynamic image preprocessing path.
• When enabled, LMDeploy skips dynamic_preprocess(...) and applies the existing model transform to the original image as a single patch.
• Document max_dynamic_patch and skip_preprocess in the English and Chinese InternVL docs.
• Add a regression test that fails if skip_preprocess=True still calls dynamic_preprocess.

BC-breaking (Optional)

No. The default preprocessing path is unchanged.

Use cases (Optional)

messages = [dict(role='user', content=[
    dict(type='text', text='Describe this image.'),
    dict(type='image_url', image_url=dict(
        url='https://example.com/image.jpg',
        max_dynamic_patch=1,
        skip_preprocess=True))
])]

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
   • PYTHONPATH= python -m pre_commit run --files lmdeploy/vl/model/internvl.py tests/test_lmdeploy/test_vl/test_internvl_preprocess.py docs/en/multi_modal/internvl.md docs/zh_cn/multi_modal/internvl.md
2. The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
   • python -m pytest tests/test_lmdeploy/test_vl/test_internvl_preprocess.py -q
3. If the modification has a dependency on downstream projects of a newer version, this PR should be tested with all supported versions of downstream projects.
   • N/A
4. The documentation has been modified accordingly, like docstring or example tutorials.
   • Updated docs/en/multi_modal/internvl.md and docs/zh_cn/multi_modal/internvl.md.

[lvhan028]

STOP pull new request until fixing the previous ones.