feat(gpt): 接入 TwelveLabs Pegasus 视频理解后端（可选）

[mohit-twelvelabs]

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

中文摘要

为 BiliNote 新增一个可选的 GPT 后端：当供应商 type 为 twelvelabs 时，用 TwelveLabs 的 Pegasus 视频理解模型直接「看」视频，从画面 + 语音两路信息生成笔记，而不只是总结转写文本。对演示、操作录屏、图表讲解类视频尤其有用——画面里常常藏着转写拿不到的信息。

• 完全可选、不改默认行为：GPTFactory.from_config 仅在 type == "twelvelabs" 时路由到新后端，其余供应商一字未改；不配置 TwelveLabs 供应商时整条链路与原来完全一致。
• API Key 从供应商配置读取，不硬编码、不入库到仓库。
• 关键文件：backend/app/gpt/twelvelabs_gpt.py（新后端）、backend/app/gpt/gpt_factory.py（路由）、backend/app/models/gpt_model.py（GPTSource 加可选 video_url）、backend/app/services/note.py（把 URL 传给 source）、backend/requirements.txt、doc/twelvelabs.md。
• 已知限制：Pegasus 需要可公开访问的视频 URL；本地文件/纯音频任务会优雅报错并提示改用常规文本模型。详见 doc/twelvelabs.md。

可在 https://twelvelabs.io 免费申请 API Key —— 有较慷慨的免费额度。

---

English

This adds an optional GPT backend that uses TwelveLabs Pegasus video understanding to generate notes directly from a video's visual + audio content, instead of summarizing only the transcript. Great for demos, screencasts, and slide/diagram-heavy videos where the screen carries information the transcript misses.

What it adds

• A new TwelveLabsGPT backend in backend/app/gpt/twelvelabs_gpt.py implementing the existing GPT.summarize interface.
• GPTFactory.from_config routes to it only when a provider's type == "twelvelabs". Every other provider path is untouched.
• GPTSource gains an optional video_url (additive, defaults None); NoteGenerator passes the source URL through. Pegasus watches the video URL directly.
• Reuses the existing note prompt (generate_base_prompt) so style/format/tag options still apply.

Why it helps this project

BiliNote already markets "多模态视频理解"; this gives users a true video-understanding path: the model sees the video rather than relying on transcription quality, which matters for visual-heavy content.

Opt-in / non-breaking

If no TwelveLabs provider is configured, behavior is identical to before. No defaults change.

How it was tested

• No-network unit tests (backend/tests/test_twelvelabs_gpt.py, stub-based to match the repo's existing test style): verify the factory routing, that summarize calls Pegasus with the video URL and max_tokens >= 512, and that a missing URL / empty key raises clearly. Run with the repo's existing pytest; existing GPT tests still pass (no regression).
• Live contract verification against the real TwelveLabs API: confirmed analyze(model_name="pegasus1.5"/"pegasus1.2", video=VideoContext_Url(...) | videoId, prompt=..., max_tokens=...) returns Markdown notes synthesized from both visual and audio tracks. Also confirmed the API's max_tokens >= 512 constraint (a 400 parameter_invalid otherwise), which the implementation enforces via _MIN_MAX_TOKENS.
• python -m py_compile passes on all changed backend files.

Known limitations (stated honestly)

• Pegasus needs a publicly reachable video URL; local-file / audio-only tasks have no URL and the backend raises a clear message asking the user to use a regular text model. (Some sample-host MP4s are rejected by Pegasus as video_file_broken due to container/codec; real platform CDN URLs from Bilibili/YouTube are fine.)
• max_tokens floor of 512; analyzed window must be ≥ 4s.

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

Note on branch target

Per CONTRIBUTING.md, features target develop. This PR is opened against develop. Happy to adjust scope/conventions per maintainer preference.