feat: 可选 TwelveLabs 视频理解 provider（Pegasus 问答 + Marengo 向量）/ opt-in TwelveLabs video understanding

[mohit-twelvelabs]

Hi! I'm Mohit, I work at TwelveLabs (@mohit-twelvelabs).

简介（中文）

为 ViNote 新增可选的 TwelveLabs 视频理解 provider，让「视频问答」可以直接理解视频画面（动作、场景、画面中的物体等转录文本无法表达的信息），而不仅仅依赖转录文本。

• Pegasus —— 直接对视频内容进行问答（TwelveLabsVideoQAService，接口与现有 VideoQAService 完全一致，可无缝替换）。
• Marengo —— 生成 512 维多模态向量，用于对抽取出的视频片段做语义检索。

完全可选、非破坏性：通过 VIDEO_QA_PROVIDER=twelvelabs + TWELVELABS_API_KEY 显式启用；不配置时，现有「转录文本 + OpenAI」路径完全不受任何影响。依赖以可选 extra 形式提供（uv sync --extra twelvelabs）。

可在 https://twelvelabs.io 免费获取 API key，有较慷慨的免费额度。

---

What this adds (English)

This PR adds an opt-in TwelveLabs video-understanding provider so ViNote's video Q&A can reason over the actual video (actions, scenes, on-screen objects — things a transcript can't capture), not just the transcript text.

• Pegasus — answers questions directly over the video. TwelveLabsVideoQAService mirrors the existing VideoQAService interface (answer_question_stream / is_available), so it slots in as a drop-in QA provider.
• Marengo — produces 512-dim multimodal embeddings for semantic search over extracted clips.

Why it helps ViNote

ViNote already does transcript-based Q&A. For visually-rich videos (tutorials, demos, sports, anything where what's on screen matters), transcript-only answers miss a lot. Pegasus closes that gap, and Marengo gives a foundation for "find the clip that matches this query" semantic search.

Opt-in / non-breaking

• Disabled by default. Only active when VIDEO_QA_PROVIDER=twelvelabs and TWELVELABS_API_KEY are set.
• With no key configured, get_video_qa_service() returns the existing VideoQAService unchanged — zero behavior change.
• twelvelabs is an optional dependency ([project.optional-dependencies] → uv sync --extra twelvelabs), so existing installs are untouched.

How it was tested

• Live-verified the TwelveLabs contracts against the real API with the official SDK (twelvelabs>=1.2.8):
  • Marengo marengo3.0 text embedding returns a 512-dim vector.
  • Pegasus pegasus1.5 analyze(...) returns a text answer (verified via the asset_id path on an uploaded clip).
• 5 pytest tests pass (backend/tests/test_twelvelabs_service.py): no-network unit tests for config toggles, provider selection, sentence-chunking, and the empty-URL guard, plus a live Marengo 512-dim contract test that auto-skips without TWELVELABS_API_KEY.

Notes / verified gotchas baked in

• Pegasus 1.5 doesn't accept a bare video_id — this uses a public URL or an uploaded asset_id (direct upload capped at 200MB, documented in code + README).
• Pegasus requires the analyzed window ≥ 4s, and max_tokens ≥ 512 (clamped in config).

You can grab a free API key at https://twelvelabs.io — there's a generous free tier.