feat(models): native response_format pass-through for OpenAI / OCIOpenAIModel

[fede-kamel]

Closes #36.

Problem

When `Agent(output_schema=Pydantic)` is configured, the agent loop always falls back to prompted-JSON + post-hoc parsing — even on OpenAI / OCI-OpenAI-compat where `response_format={"type":"json_schema",...}` is supported natively.

The native pass-through path existed in the codebase but was only used in the post-loop repair flow at `agent.py:1743`, never in the main loop at `agent.py:1632`. `build_response_format()` (`src/locus/core/structured.py:257-288`) already returns the right shape; it just wasn't being called.

Fix

Add a `supports_structured_output` capability property on each provider model:

Provider	Returns	Reason
`OpenAIModel`	`True`	Native `response_format`
`OCIOpenAIModel`	`True` (inherited)	Same chat-completions shape on OCI
`AnthropicModel`	`False`	No `response_format`; prompted-JSON path
`OllamaModel`	`False`	Same
`OCIModel` (Cohere R-series via native SDK)	`False`	Different transport

In `agent.py`, the main-loop `complete()` call now gates on the capability:

```python
if self.config.output_schema is not None and getattr(
self._model, "supports_structured_output", False
):
response_format = build_response_format(
self.config.output_schema,
strict=self.config.output_schema_strict,
)
```

When set, `response_format=` is passed to `model.complete(...)` in the same kwargs path the existing repair flow already uses.

Outcome

• OpenAI / OCI-OpenAI-compat: native pass-through. The provider returns a guaranteed-parseable Pydantic instance — no retry, lower token cost, stronger correctness guarantee.
• Other providers: unchanged. Prompted-JSON fallback still active. Backward-compatible.

Tests

Six new tests in `tests/unit/test_native_structured_output.py`:

• Each provider's `supports_structured_output` value verified.
• `build_response_format(SamplePayload, strict=True)` returns the OpenAI shape.

All 3,210 pre-existing unit tests still pass; pre-commit clean; `mkdocs build --strict` clean.

Test plan

• `hatch run test tests/unit/` — 3,210 pass + 1 skip (OCIOpenAIModel needs real config)
• `pre-commit run --all-files` — clean
• `hatch -e docs run mkdocs build --strict` — clean
• DCO sign-off