fix(pydantic_ai): report structured token usage instead of stringifying to evaluation_context (#28)

[mhaye9545]

Closes #28.

Summary

• Extracts structured \input_tokens, \output_tokens, and \ otal_tokens\ from PydanticAI's
  esult.usage()\ and passes them to \Result(..., usage=...)\ instead of stringifying them into \evaluation_context.
• Regenerated the \pydantic_ai_calculator.json\ conformance fixture via \scripts/seed_adapter_fixtures.py.
• Added unit tests in \ est_adapters.py\ to verify structured usage extraction and clean evaluation context.