Add text_image modality support for text elements in multimodal embedder

[edknv]

Description

This PR enables the ingestion pipeline to pair extracted text with its source image for multimodal embedding (e.g., using llama-3.2-nemoretriever-1b-vlm-embed-v1). It also adds logic to aggregate all text content from a PDF page and attach it to a full-page image, which yields the best recall on multiple datasets (bo10k, earnings). Despite the improved recall of this path, it is not set as default due to lower throughput.

• Added nvidia/llama-nemotron-embed-vl-1b-v2 to the list of supported multimodal models.
• The OCR extractor now preserves the original base64 image in metadata (as source_image) before it gets overwritten by text, allowing the embedding stage to access the visual context.
• Implemented _aggregate_page_content which collects all text from TEXT and STRUCTURED (tables/charts) elements on a specific page and attaches it to the PAGE_IMAGE entry.
• Added flags (embed_text_elements, etc.) to allow users to skip embedding individual text snippets if they are already being embedded as part of an aggregated page image.
• Added logic to automatically enable page aggregation if text_image modality is requested and page images are present.
• Added logic to detect if a base64 string is JPEG or PNG based on magic bytes, ensuring correctly formatted data URIs for embedding models.

Usage

ingestor = (
    Ingestor(message_client_hostname="nv-ingest-ms-runtime")
    .files(docs)
    .extract(
        extract_text=True,
        extract_tables=True,
        extract_charts=True,
        extract_infographics=True,
        extract_images=False,
        text_depth="page",
        table_output_format="markdown",
        extract_page_as_image=True,
    )
    .embed(
        image_elements_modality="text_image",
    )

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.
• If adjusting docker-compose.yaml environment variables have you ensured those are mimicked in the Helm values.yaml file.