feat: add litellm response_format example

[rlippmann]

What changed

• add a new direct LiteLLM SDK example that uses saved Context Compiler state to select LiteLLM response_format
• add deterministic tests for response-format planning and LiteLLM call-shape behavior
• add import coverage for the new integration example
• document the application-layer boundary for this example and clarify the local Ollama smoke command requirements

Why

• existing LiteLLM examples showed prompt/message construction from saved state, but not request-shape selection
• this example demonstrates a different and useful boundary: Context Compiler owns authoritative state, the host maps that state into LiteLLM request shape, and LiteLLM handles model invocation
• the Ollama smoke command note is needed because this repo still requires PROVIDER=ollama for local Ollama runs even when MODEL=ollama/... is set

Checklist

• pre-commit run (uv run pre-commit run --all-files)
• tests pass (uv run pytest)