fix(medhelm): score PubMedQA with punctuation-tolerant metric

[MukundaKatta]

Summary

• compute quasi_leave_articles_exact_match for the MedHELM PubMedQA run spec
• add the metric to the MedHELM schema
• use it as the PubMedQA main metric so answers like "A." are scored consistently

Fixes #4187.

Tests

• python3 -m compileall -q src/helm/benchmark/run_specs/medhelm_run_specs.py src/helm/benchmark/metrics/evaluate_reference_metrics.py
• python3 schema assertion for quasi_leave_articles_exact_match and PubMedQA main metric
• git diff --check

[goktugozkanmd]

One small consistency check before merge: the PR changes the MedHELM schema so PubMedQA reports quasi_leave_articles_exact_match as the main metric, but PubMedQAScenario.get_metadata() still declares main_metric="exact_match".

If that metadata is only legacy display data, no issue. If any docs, registry output, or downstream summary path reads it, users could still see PubMedQA described as exact match while the MedHELM schema is scoring the punctuation tolerant metric.

I would update the scenario metadata too, or leave a short note that schema_medhelm is the only source used for this leaderboard row.