Summary

This PR fixes a provider-specific request mapping issue for Vertex AI and adds the required Vertex AI SDK
extra in dependencies.

When using vertexai:* models, Bub previously passed extra_headers into the LLM call. In the Vertex AI
path, that value is validated against GenerateContentConfig, where extra_headers is invalid, causing
runtime errors.

What Changed

• Updated ModelRunner request construction:
• For vertexai:* models, pass headers via http_options={"headers": ...}.
• For non-Vertex providers, keep the default behavior and pass extra_headers.
• Added regression coverage in tests/test_model_runner.py:
• OpenRouter/non-Vertex path still uses extra_headers.
• Vertex AI path uses http_options.headers and does not send extra_headers.
• Unknown providers continue using default extra_headers.
• Updated dependencies:
• any-llm-sdk[anthropic] -> any-llm-sdk[anthropic,vertexai]
• Synced lockfile accordingly.

Why

This PR fixes a provider-parameter mismatch that is reproducible with Vertex AI:

• Set BUB_MODEL=vertexai:gemini-3.1-pro-preview
• Run uv run bub and send a simple message
• Before this change, Bub passed extra_headers in the model call
• In the Vertex AI path, forwarded kwargs are validated against GenerateContentConfig, where
  extra_headers is invalid
• This caused runtime failure: GenerateContentConfig ... extra_headers ... Extra inputs are not permitted

The fix maps headers correctly for Vertex AI (http_options.headers) while keeping default extra_headers
behavior for non-Vertex providers.

Validation

• uv run pytest -q tests/test_model_runner.py (11 passed)

• uv run ruff check src/bub/core/model_runner.py tests/test_model_runner.py (passed)

• Fixes Vertex AI runtime failure (extra_headers validation error).

• Preserves existing behavior for non-Vertex providers.

• No public API changes.