feat(ai): add vision support for multimodal messages

[willgriffin]

Summary

Add support for vision-capable LLMs by extending AIMessage.content to accept both simple strings and multimodal content arrays (ContentPart[]).

Changes:

• Add ContentPart types (TextContentPart, ImageContentPart) for multimodal messages
• Add extractTextContent() helper function for backward compatibility
• Update AIMessage.content type to accept string | ContentPart[]
• Update OpenAI provider to properly map ContentPart arrays to OpenAI format
• Update all other providers (Anthropic, Bedrock, Claude-CLI, Gemini, HuggingFace) to use extractTextContent() for backward compatibility

Usage

This enables vision-capable models to accept images inline:

import { getAI } from '@happyvertical/ai';

const ai = await getAI({ type: 'openai' });

const response = await ai.chat([
  {
    role: 'user',
    content: [
      { type: 'text', text: 'What is in this image?' },
      { 
        type: 'image_url', 
        image_url: { 
          url: 'data:image/png;base64,iVBORw0KGgo...', 
          detail: 'high' 
        } 
      }
    ]
  }
]);

Motivation

Required for @happyvertical/ocr LiteLLM provider which uses vision-capable LLMs for OCR text extraction.

See: happyvertical/ocr#21

Test plan

• All existing tests pass (71 tests, 10 skipped for API keys)
• TypeScript compilation passes
• Manual testing with vision-capable model (GPT-4o, DeepSeek-VL)

[github-actions]

📦 Version Bump Preview

When this PR is merged, packages will receive a minor
version bump based on your conventional commits.

What happens on merge?

1. Tests run on main branch
2. Packages are built
3. Versions are bumped automatically
4. Packages are published to GitHub Packages
5. Git tags are created

No manual intervention needed! 🎉