feat: add TwelveLabs Pegasus as a video asset-analysis provider

[mohit-twelvelabs]

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

中文摘要

本 PR 新增 TwelveLabs Pegasus 作为可选的视频素材分析（VLM）后端，与现有的 DashScope/Qwen 并列。Pegasus 是视频原生理解模型，会理解整段视频（运动、场景切换、动作），而非仅采样若干帧，因此非常适合在生成脚本/分镜前理解用户上传的视频素材。

• 完全可选、不改变默认行为：未配置 twelvelabs API Key 时行为不变。
• 在「素材分析服务」中作为仅视频选项出现（Pegasus 不分析图片）。
• 已对 TwelveLabs 真实 API 做了端到端验证（上传 → 资产就绪轮询 → 分析返回文本），并附带单元测试。

可在 https://twelvelabs.io 免费获取 API Key（提供慷慨的免费额度）。

What this adds

This PR adds TwelveLabs Pegasus as an opt-in video asset-analysis (VLM) provider, alongside the existing DashScope/Qwen provider used by APIAssetAnalysisService. Pegasus is a video-native understanding model — it reasons over the whole clip (motion, scene changes, on-screen action) rather than a handful of sampled frames — which makes it a strong fit for understanding uploaded footage before script/storyboard generation.

Concretely:

• PegasusClient (vlm_twelvelabs.py) mirrors the existing vlm_dashscope.py client style. It accepts public URLs and local files (local files are uploaded as a TwelveLabs asset; direct upload is capped at 200MB, larger footage should be passed as a public URL).
• The unified VLM client routes pegasus* models to TwelveLabs and everything else to DashScope. Pegasus is video-only and cleanly rejects image input.
• Wired into APIAssetAnalysisService.list_models() (exposed as a video-only option) and the api_providers config schema + config.example.yaml.
• Added twelvelabs>=1.2.8 (the official SDK) to dependencies.

Why it helps this project

Pixelle-Video already supports swapping analysis backends (RunningHub / self-host ComfyUI / direct API VLM). For users who feed real video clips into the asset-based pipeline, a video-native model produces noticeably more grounded descriptions than frame-sampling VLMs, which improves downstream script and storyboard quality. It slots into the existing provider/registry pattern, so it's just one more option in the dropdown.

Opt-in & non-breaking

If no twelvelabs API key is configured, nothing changes — the provider is simply not listed. No defaults are altered.

How it was tested

• Live-verified against the real TwelveLabs API in this PR's development: the full PegasusClient path (local file upload → asset-ready poll → analyze) returns a correct text description; I also confirmed the API contract directly.
• Unit tests cover provider routing, the configured-only listing, the video-only image-rejection guard, and the "requires video" guard (no network). A live test (skipped unless TWELVELABS_API_KEY is set) exercises the real upload→analyze path. Note: the repo's .gitignore excludes test_*.py, so the test file is not committed — happy to add it under a different path/name if you'd like it tracked.
• ruff check passes clean on the new files.

Possible follow-up

TwelveLabs also offers Marengo multimodal embeddings (512-dim), which would enable footage retrieval/selection (ranking uploaded clips against a query) — a natural next step if you'd like a selection backend; happy to follow up in a separate PR.

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

---

Note: TwelveLabs employee contribution. If AIDC-AI requires a CLA or maintainer sign-off for external contributions, just let me know and I'll complete it.

[CLAassistant]

All committers have signed the CLA.