Align dashboard AUC with leaderboard methodology

[mitre88]

Summary

• move AUC calculation into a small tested helper
• calculate dashboard AUC as the paper-described left Riemann sum starting at op=2
• stop accumulating once accuracy falls below 5% and use actual op spacing for non-unit strides
• add focused regression tests for start-op, truncation, stride, and sorting behavior

Validation

• python -m unittest discover -s tests -v
• python -m py_compile gsm-infinite/metrics.py gsm-infinite/app.py tests/test_metrics.py
• git diff --check

Context: issue #9 asks about the missing AUC truncation / op-stride behavior. This PR implements that dashboard-side behavior without touching the existing preprocess or answer-parsing PRs.

/claim https://algora.io/PrimeIntellect-ai/bounties/m4592Ap5jB1qm9yd

[mitre88]

/claim https://algora.io/PrimeIntellect-ai/bounties/m4592Ap5jB1qm9yd

Claiming after submitting the implemented and validated dashboard AUC fix. Validation is listed in the PR body.