fix multi-modality apply chat template issue

[sywangyi]

What does this PR do?

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[sywangyi]

@regisss @Narsil please help review

[regisss]

@sywangyi Can you add a description of the issue and how this PR solves it please?

[sywangyi]

#3257 describe it, And I just give a WA as skipping applying chat template to VLM model like llava qwenvl

[regisss]

Gently pinging @Narsil for review

[Narsil]

This is most likely a wrong fix. We cannot just skip using templates on some models because the answer is not to our liking.

We need to fix the template instead, no ? But overriding like this is most likely purely wrong.

[sywangyi]

VLM template input like,
conversation = [
{

  "role": "user",
  "content": [
      {"type": "text", "text": "What is shown in this image?"},
      {"type": "image"},
    ],
},

]

which is different with llm.
{"role": "user", "content": "Who are you?"},

so we may need reconstruct the msg for vlm before apply chat template. also, there is model that does not have chat template. we need to handle it as well.