evaluator: validate persisted runs before resuming

[huangyz0918]

Summary

• Validate CSV dataset identity, value types, and row order before reusing predictions.
• Fingerprint checkpoints by dataset, predictor model, config, and skip policy while retaining legacy-checkpoint compatibility.
• Preserve structured and mixed scalar values across CSV round trips.
• Preserve Boolean comparator scores, legacy JSON key representations, and exact-value legacy pandas numeric CSV formatting.
• Report all-skipped and empty evaluations as N/A.
• Accept numeric SVAMP JSON answers.

Why

A same-length reordered dataset or stale checkpoint could previously attach predictions to the wrong rows. Resumed runs could also mix scoring policies, and all-skipped runs could crash while formatting accuracy.

Verification

• Unchanged repository suite: 123 passed, including PyTorch paths.
• Additional local regression checks cover reordered datasets, CSV round trips, checkpoint identity and compatibility, empty/skipped results, and numeric SVAMP answers.
• Only production bug fixes are included. Additional validation code stays local; no credentials or live API calls are involved.