fix: send extra headers only for openrouter provider

[Hopsken]

Summary

Fix provider-specific header routing in model calls to avoid passing OpenRouter metadata to incompatible providers (for example Gemini).

What Changed

• Updated ModelRunner._chat header logic:
  • openrouter:* -> pass extra_headers (HTTP-Referer, X-Title)
  • vertexai:* -> pass http_options.headers (existing behavior)
  • all other providers -> pass no provider-specific headers
• Updated tests in tests/test_model_runner.py to match the stricter routing behavior.

Why

Gemini rejects unknown config fields like extra_headers, causing runtime validation errors.
Routing headers only where needed prevents cross-provider incompatibilities.

Verification

• uv run pytest -q tests/test_model_runner.py
• Result: 12 passed

Impact

• Fixes Gemini runtime error with BUB_MODEL=gemini:...
• Keeps OpenRouter attribution headers for OpenRouter
• Preserves VertexAI-specific header path

[Hopsken]

Just notice that there is a big ongoing refactoring, closing this to avoid conflicts.