Add opt-in strict JSON Schema output for OpenAI models, closes #1661

[wellparth]

Closes #1661.

Adds an opt-in strict_schema model option for OpenAI models that enables Structured Outputs with strict: true:

llm -m gpt-4o -o strict_schema 1 --schema 'name, age int' 'invent a cool dog'

model.prompt(prompt, schema=OutputModel, strict_schema=True)

What changed

• New strict_schema option available on all OpenAI model options classes. Defaults to false so existing behaviour is unchanged. Setting it without a schema raises a clear error.
• Chat Completions path (_Shared.build_kwargs): adds "strict": true to the json_schema response_format.
• Responses API path (_SharedResponses._build_responses_kwargs): sets "strict": true instead of the previous hardcoded false.
• New llm/default_plugins/openai_strict_schema.py: converts a JSON schema (e.g. one produced by a Pydantic model via model_json_schema()) into the restricted strict-compatible form:
  • adds additionalProperties: false to every object,
  • adds every property to required,
  • inlines $defs/$ref definitions,
  • preserves Optional-style anyOf: [T, {type: null}] unions,
  • raises a clear UnsupportedStrictSchemaError for keywords OpenAI strict mode cannot enforce (minItems, maxItems, allOf, multipleOf, etc.), rather than silently weakening the schema.

Notes on response handling

The issue also mentions surfacing finish_reason / refusal / truncation status. That's a larger change to llm.Response and is intentionally left out of this PR to keep it focused; the compiler/migration pieces here are the prerequisite for it.

Tests

• tests/test_openai_strict_schema.py — converter unit tests (19): strict-shape normalization, $defs inlining, null unions, cosmetic-key stripping, error cases, input immutability.
• tests/test_cli_openai_models.py — 6 integration tests using httpx2_mock: chat + responses bodies with/without the option, the requires-schema error, and the unsupported-keyword error.
• Full suite: 1148 passed. ruff check and mypy clean.