Add suite-level reasoning control for provider/model commissions

[danballance]

Motivation

• Benchmarks always used Pi's maximum thinking level (xhigh) for provider models, preventing experiments with low or no reasoning; suites need per-provider control to benchmark different reasoning levels.
• The change exposes a simple, explicit reasoning configuration on provider/model commissions so reasoning can be varied per-suite without altering providers or model catalog entries.

Description

• Added a reasoning field to ProviderConfig (defaulting to "xhigh") in park_bench/models/suite.py so provider entries carry a reasoning level.
• Updated API request handling to use typed provider configs via ProviderConfig in park_bench/api/schemas/requests.py and serialize provider entries with model_dump() in park_bench/api/routes/suites.py so reasoning is persisted in suite config.
• Threaded the reasoning value through execution: SuiteRunner passes provider_cfg.reasoning into Orchestrator, Orchestrator forwards reasoning into executor calls, and BaseContainerExecutor includes --reasoning in the container entrypoint command.
• Plumbed reasoning into the container entrypoint surface by adding reasoning to IterationEntrypointArgs and EntrypointArgumentParser, and passed it into PiLLMClient; PiLLMClient now uses the supplied reasoning for Pi's --thinking argument instead of a hardcoded xhigh.
• Updated tests to assert the new command wiring and defaults across suite API, executor command construction, and orchestrator behavior.

Testing

• Ran formatting and linting with uv run ruff format park_bench tests and uv run ruff check --fix park_bench tests with no unresolved issues.
• Ran complexity validation with uv run python .github/scripts/check_complexity.py (no failures).
• Ran the full test suite with uv run pytest --cov park_bench --cov-report term-missing tests and all tests passed (991 passed, 5 warnings).

---

Codex Task