feat: Add pii_detection check

[harsh21234i]

#2374

Summary

Adds a new built-in pii_detection check to giskard-checks for
detecting personally identifiable information in model outputs.

What changed

• added PIIDetection under giskard.checks.judges
• added prompt template judges/pii_detection.j2
• exported the new check from public package entrypoints
• added tests for:
  • structured PII detection in pattern mode
  • contextual PII detection in llm mode
  • hybrid mode short-circuiting on regex hits
  • category filtering
  • serialization

Behavior

Supported modes:

• pattern
• llm
• hybrid

Structured categories currently handled with regex:

• email
• phone
• ssn
• credit_card
• ip_address

Contextual categories handled through LLM evaluation:

• name
• address
• medical_info
• financial_info

hybrid mode first checks structured PII via regex and only falls back to
the LLM for contextual categories when needed.

Validation

Ran:

uv run --all-packages --group dev pytest libs/giskard-checks/tests/
builtin/test_pii_detection.py libs/giskard-checks/tests/builtin/
test_toxicity.py

Result:

- 20 passed

## Related

Closes #2374

[gemini-code-assist]

Code Review

This pull request introduces the PIIDetection check, which identifies personally identifiable information in AI responses using regex patterns, LLM evaluation, or a hybrid approach. Feedback focuses on refining the regex patterns for SSNs, credit cards, and IP addresses to reduce false positives, as well as correcting a logic error in the output resolution process that could cause runtime failures when using default settings.

[harsh21234i]

Addressed the review feedback by fixing trace fallback for output extraction and tightening the SSN, credit-card, and IPv4 regexes. Added tests for the stricter pattern behavior and reran the targeted giskard- checks suite successfully.
Happy To Contribute More !!!