fix(server): keep OpenAIChatMessage validator importable across the supported pydantic range

[feiiiiii5]

Summary

Fixes #2292.

OpenAIChatMessage passed json_schema_input_type to BeforeValidator, but that argument is only accepted from pydantic 2.9 onwards, while the project declares pydantic>=2.5,<3.0. Importing nemoguardrails.server.schemas.openai on pydantic 2.5–2.8 therefore raised:

TypeError: BeforeValidator.__init__() got an unexpected keyword argument 'json_schema_input_type'

This change builds the validator through a small helper that applies the argument when the installed pydantic supports it and falls back to a plain BeforeValidator otherwise:

def _build_chat_message_validator() -> BeforeValidator:
    try:
        return BeforeValidator(
            _validate_openai_chat_message,
            json_schema_input_type=_OpenAIChatMessageInput,
        )
    except TypeError:
        return BeforeValidator(_validate_openai_chat_message)

On pydantic < 2.9 the generated JSON schema describes messages as generic objects instead of the per-role union; runtime validation behavior (string role required, per-role payload validation) is identical.

I kept the declared dependency range as-is rather than bumping the minimum to 2.9, per the issue's "either remove/replace the argument or raise the minimum" — happy to switch to a floor bump + lockfile regeneration if maintainers prefer that route.

Verification

• Reproduced the original TypeError empirically on pydantic 2.5.1, 2.6.4, 2.7.4 and 2.8.2 before the change; confirmed 2.9.2 accepts the argument.
• New tests in tests/server/test_openai_schemas_compat.py:
  • valid user message round-trips through validation,
  • unknown role is rejected,
  • JSON schema generation succeeds on every supported version,
  • the rich per-role input type is asserted when pydantic supports it (auto-skipped below 2.9).
• Ran the new tests against pydantic 2.8.2 (fallback path: 3 passed, rich-schema test skipped) and 2.13.4 (rich path: 4 passed).
• ruff check and ruff format --check clean with the repo-pinned ruff 0.14.6.

AI Disclosure

Per AI_POLICY.md: this pull request was written with AI assistance (ox-alpha coding agent operating on behalf of the submitting human) and reviewed and edited by a human, who verified the changes locally across the pydantic versions listed above. No AI co-author is recorded; the commit is DCO signed by the submitter.

[feiiiiii5]

Closing as part of a contribution-portfolio consolidation: my effort is now focused on a single project (failroute, a static analyzer for silent-failure anti-patterns) and its upstream follow-through. Happy to revisit if there is interest.