Typed video input support: VideoURL multimodal class

[Eigenwise]

PR #274 adds video input support by detecting raw {"type": "video_url", "video_url": {"url": ...}} dicts in schema fields. The use case is real (MiniMax-M3 and other OpenAI-compatible providers accept video_url content parts), but the implementation has a blocker: LiteLLM's token_counter raises on video_url parts, so any agent with max_context_tokens set crashes on every run() once a video is in history. Shape-detection of bare dicts also cuts against the schema-first design. Details in the review on #274.

Plan:

• Add a typed VideoURL class (url, optional fps / detail, to_openai() emitting the OpenAI-compatible content part), exported from the package root
• Treat it as a first-class multimodal type in ChatHistory._extract_multimodal_info, same as Instructor's Image / Audio / PDF
• get_history() emits the content-part dict (Instructor passes plain dicts through to the provider unchanged, verified against 1.14.x)
• Token counting uses a text placeholder for video parts since LiteLLM can't count them
• dump() / load() round-trips it natively as a plain Pydantic model

Instructor has no Video type of its own yet; upstream ask is 567-labs/instructor#2520. If that lands we can swap it in behind the same seam.