fix(fns): handle BinaryImage/UserContent types as attachments in extract() and classify()

[octo-patch]

Summary

Extends the UserContent attachment handling (similar to #1257 for cast()) to extract() and classify().

Problem

When passing pydantic-ai UserContent types (BinaryImage, ImageUrl, AudioUrl, DocumentUrl, VideoUrl) to marvin.extract() or marvin.classify(), the binary data is serialized into the task context as a string, causing token explosion (200K+ tokens for moderately-sized images).

Example that fails with token explosion:

from pathlib import Path
from pydantic_ai.messages import BinaryImage
import marvin

image_bytes = Path("image.png").read_bytes()
binary_image = BinaryImage(data=image_bytes, media_type="image/png")

# Token explosion: binary data serialized into context
tags = marvin.extract(binary_image, target=str, instructions="List objects in the image")
label = marvin.classify(binary_image, labels=["safe", "unsafe"])

Solution

Added detection for pydantic-ai's UserContent types in extract() and classify() — the same fix applied to cast() in #1257. Binary/URL content types are now passed as task attachments instead of being serialized into the context dictionary.

Testing

• Added unit tests for extract() with BinaryImage and ImageUrl inputs
• Added unit tests for classify() with BinaryImage and ImageUrl inputs
• All 6 new tests pass