feat(dataset): add generative CivilComments config

[motodriver]

Motivation

The existing CivilComments configuration uses CLPInferencer and AUCROCEvaluator, which require class logits. Text-only API models cannot run that path.

Changes

• keep the existing CLP/AUC configuration unchanged
• add a separate civilcomments_gen configuration using GenInferencer and AccEvaluator
• map non-toxic comments to A/No and toxic comments to B/Yes
• parse only A/B answers with cushion=False
• expose both stable configurations in dataset-index.yml
• add offline tests for the 0.5 threshold boundary, label direction, config structure, and CLP compatibility

Verification

• focused offline tests: 4 passed
• YAPF: passed
• Python compilation: passed
• dataset suffix/hash verification: passed
• git diff --check: passed

Compatibility

The original CLP configuration and AUC metric remain available. The new generated-answer path reports accuracy and does not attempt to emulate probability-based AUC.

Closes #1416.