Promote `TwelveLabsVideoParser` and `MarengoEmbedder` from the Video RAG template into `pathway.xpacks.llm`

[zxqfd555]

Background

pathwaycom/llm-app#129 added a Video RAG with TwelveLabs template containing two reusable building blocks:

• TwelveLabsVideoParser (pw.UDF) — uploads video bytes as a TwelveLabs asset and turns them into a rich text description via the Pegasus video-understanding model. Output is the standard list[(text, metadata)], so it chunks/embeds/indexes exactly like the built-in PDF parsers.
• MarengoEmbedder (BaseEmbedder) — a retriever embedder backed by the Marengo multimodal model, returning 512-dim vectors in a shared text/image/audio/video embedding space.

Both currently live in a per-template pathway_twelvelabs package. This issue proposes promoting them into pathway.xpacks.llm as first-class, supported components.

Motivation

• Fills a real gap in modality coverage. xpacks/llm/parsers.py already ships ImageParser and AudioParser but has no video parser. MarengoEmbedder would be the first xpacks embedder in a shared multimodal space. New modality, zero new infrastructure.
• Already conforms to xpacks conventions. MarengoEmbedder subclasses BaseEmbedder like OpenAIEmbedder/GeminiEmbedder/BedrockEmbedder and reuses udfs.async_executor, ExponentialBackoffRetryStrategy, CacheStrategy, max_batch_size. TwelveLabsVideoParser mirrors the existing pw.UDF parsers. Minimal adaptation required.
• Generic, not template-specific. Prompt, model, API key, retries, and caching are all parameterized; no template-specific assumptions in the classes themselves.

Proposed scope

• Add the two components under pathway/xpacks/llm/ fold into embedders.py / parsers.py with exports in the respective __init__.
• Keep twelvelabs an optional dependency via the existing lazy-import + ImportError guidance pattern, so plain import pathway is unaffected. Create a twelvelabs section in pyproject.toml so that the dependency can be installed via pip install pathway[twelvelabs].
• Port the unit tests (stubbed SDK, no network) into python/pathway/xpacks/llm/tests/; keep the live smoke test gated behind TWELVELABS_API_KEY.
• Ensure all introduced classes are properly documented.
• Add a CHANGELOG entry (user-visible Python API).

Shortly after the components are ported into the Pathway framework and the changes are released, it makes sense to refactor the llm-app template so that it imports components directly from Pathway.

[iapoorv01]

Hi @zxqfd555 ! I'd love to take this one on if it's available. I'll get started on porting the TwelveLabsVideoParser and MarengoEmbedder over to xpacks.llm along with the optional twelvelabs dependency.