Fix ICI pooling precedence and inherited regimen selection

[iskandr]

Summary

• Prefer direct evidence for the requested endpoint and regimen before source-scope fallback in pooled_ici_response(), restoring ADCC trial points and confidence intervals in the forest plot.
• Exclude non-poolable context before source selection, and keep verification/primary-only filters from silently switching populations.
• Make selected_ici_regimen() use the same cached, inheritance-aware resolver as the response lookup, with an explicit inherit option. Preserve the legacy cancer_ici_regimen() behavior.
• Correct the regimen tags and pooling eligibility of three existing salivary contextual estimate rows. Numeric values, counts, citations, and estimate IDs are unchanged; no data rows are added or removed.
• Document the behavior and add 17 regression cases covering direct/source-scoped lookup, endpoint and filter handling, whole-registry regimen parity, and the ADCC forest plot.

The contextual-row corrections are supported by Vos et al. and KEYNOTE-158.

Validation

• ./test.sh: 1,221 passed, Ruff lint and formatting clean.
• pandas 3.0.5: 127 focused evidence/API tests passed, plus the ADCC forest-plot regression.
• Confirmed the new tests reproduce the original defects before applying the fixes.
• git diff --check: clean.