experimental: introduce img understand pipeline

[dolfim-ibm]

This new feature creates a new ImgUnderstand pipeline which uses vision LLMs to describe the pictures contained in documents.

The pipeline allows to use

1. Local LLM, via vLLM
2. LLM as a service, e.g. on watsonx.ai or openai compatible apis

Checklist:

• Commit Message Formatting: Commit titles and messages follow guidelines in the
  conventional commits.
• Documentation has been updated, if necessary.
• Examples have been added, if necessary.
• Tests have been added, if necessary.

[dolfim-ibm]

Offline LLM

vLLM

Pros:

• efficiently run vision models offline, see the docs page.
• supports different models without further specialization
• already used by InstructLab and part of RHEL AI

Cons:

• no support for mac (any architecture)
• vLLM has an exact pinning of torch, which creates issues with poetry.
  • vllm==0.5.x depends on torch==2.3.0
  • vllm==0.6.x depends on torch==2.4.0.

HF transforms

Pros:

• no strong pinning of torch

Cons:

• more code needed
• different models require different implementations, e.g. llava-next is different than phi-3-v.

[cau-git]

Let's call that min_area_frac to be aligned.

[PeterStaar-IBM]

@dolfim-ibm could we not use some standard HF models (eg florence and onechart)?