feat(ocr): add LiteLLM provider for vision LLM-based OCR

[willgriffin]

Summary

Add a new OCR provider that uses vision-capable LLMs (like DeepSeek, GPT-4o) via LiteLLM-compatible API endpoints. This enables OCR through any OpenAI-compatible vision model API.

Features:

• Support for both simple (text-only) and structured (JSON with confidence) output modes
• Environment variable configuration (HAVE_OCR_LITELLM_*)
• Constructor options with precedence over environment variables
• Automatic image format detection (PNG, JPEG, GIF, WebP, BMP)
• Base64 encoding for API transmission
• Comprehensive language support via vision LLM capabilities (60+ languages)

Configuration:

• HAVE_OCR_LITELLM_BASE_URL: API endpoint (default: http://localhost:4000/v1)
• HAVE_OCR_LITELLM_API_KEY: API key for authentication
• HAVE_OCR_LITELLM_MODEL: Model name (default: deepseek-chat)
• HAVE_OCR_LITELLM_OUTPUT_MODE: Output mode (simple | structured)

Provider priority: onnx > tesseract > litellm (API-based providers last)

Changes

• New files:
  • src/node/litellm.ts - LiteLLM provider implementation (~500 lines)
  • src/node/litellm.spec.ts - Comprehensive unit tests (25 tests)
• Modified files:
  • src/shared/factory.ts - Added provider loading and priority
  • src/shared/types.ts - Added LiteLLMProviderOptions interface
  • src/index.ts - Added exports for new provider
  • package.json - Added @happyvertical/ai dependency

Notes

• Uses local ContentPart type definitions since @happyvertical/ai doesn't export vision types yet
• The AI package will need a separate PR to add proper vision support
• Uses type assertion (contentParts as any) for compatibility until AI package is updated

Test plan

• All existing tests pass (27 tests)
• New unit tests pass (25 tests)
• Total: 52 tests passing
• Manual testing with actual LiteLLM/DeepSeek endpoint