Refactor tool schemas to use pydantic-ai toolsets

[cancelself]

Follow-up from PR #1 review by @milani.

Context

src/rules_agent/rules/eval.py:105-107 (and similar sites like src/rules_agent/rules/extractor.py) hand-roll OpenAI-compatible tool schemas as dict literals. pydantic-ai tools/toolsets would let us define these as typed Python objects and make the code more idiomatic.

Scope

• Evaluate whether pydantic-ai plays nicely with the existing OpenAI-SDK-based multi-provider setup (Ollama, OpenAI, Anthropic).
• Migrate the eval judge tool schema and the extractor tool schema to pydantic-ai toolsets.
• Update tests in tests/test_eval.py and tests/test_extractor.py to match.

Notes

• Non-blocking — suggestion was approved alongside PR Migrate Poetry → uv, black → ruff, fix multi-provider support #1.
• Larger refactor; touches the core LLM-call path, so worth doing as a focused PR.

Original review comment: #1 (comment)