[Feature Request] Amazon Bedrock Knowledge Bases now processes multimodal data

## Describe the solution you'd like

Support `Amazon Bedrock Knowledge Bases now processes multimodal data`
https://aws.amazon.com/jp/about-aws/whats-new/2024/12/amazon-bedrock-knowledge-bases-processes-multimodal-data/

## Why the solution needed

It will be able to generate answers with high accuracy from PDFs containing tables, etc. that it have not been good at until now.
When a PDF is vectorized, an image for each PDF page is stored in S3, and the S3 URI of the text extracted from the image and the S3 URI of the image are stored in a vector DB as a set.
If the search results in text extracted from the image, the text and image are passed together to LLM for an answer.
By handing over not only text but also images to the LLM, it is possible to respond with high accuracy even with complicated tables.

## Additional context

Add any other context or screenshots about the feature request here.

## Implementation feasibility

Are you willing to collaborate with us to discuss the solution, decide on the approach, and assist with the implementation?

- [X] Yes, I am able to implement the feature and create a pull request.
- [ ] No, I am unable to implement the feature, but I am open to discussing the solution.

Start by reviewing the linked Amazon Bedrock announcement and the repository's existing Knowledge Bases, PDF ingestion, S3, vector database, and LLM integration entry points. Define the implementation scope around preserving page images with extracted text and passing both to the model for multimodal answers, then add coverage for PDFs containing tables.

Written by the indexing model from the issue text.