Add typed VideoURL multimodal support

[Eigenwise]

Typed video input support, superseding the approach in #274 (thanks @octo-patch for kicking this off and for the MiniMax use case).

Instructor has no video type (asked upstream: 567-labs/instructor#2520), so this adds our own VideoURL:

from atomic_agents import BaseIOSchema, VideoURL
from pydantic import Field

class VideoAnalysisInput(BaseIOSchema):
    """Input with a video for analysis"""
    question: str = Field(..., description="Question about the video")
    video: VideoURL = Field(..., description="Video to analyze")

• ChatHistory treats VideoURL as a first-class multimodal type; get_history() emits it as an OpenAI-compatible video_url content part, which Instructor forwards to the provider unchanged
• Token counting turns video parts into a [video content] text placeholder. LiteLLM's token_counter raises on video_url parts, which is what broke fix: preserve video content parts in chat history #274: get_context_token_count() crashed whenever max_context_tokens was set
• Round-trips through dump() / load() as a plain Pydantic model
• Docs section in the memory guide

Closes #275. When Instructor grows a native Video type we can swap it in behind the same seam.

Checks: uv run black (unchanged), uv run flake8 (clean), uv run pytest --cov=atomic_agents atomic-agents (333 passed; test_get_class_from_string also fails locally for me on clean main, pre-existing rootdir quirk, unrelated).

🤖 Generated with Claude Code