fix(eval): classify PrefEval judge answers like the reference benchmark

[jiangwu25]

Summary

PrefEval can treat a failed judge answer as Personalized Response.

This PR fixes the classifier to match the reference behavior:

• invalid or missing answers become Judge Failure
• Yes and No checks ignore case and extra punctuation
• Unhelpful Response takes priority
• all 16 standard answer patterns match the reference logic

Tests

• 35 focused tests pass
• Ruff, format checks, pre-commit, AST checks, and git diff --check pass

Fixes #2396