[codex] Handle missing IC50 values

[iskandr]

Summary

Fixes #306.

This PR makes missing affinity IC50 values non-fatal in the two places that assumed every pMHC_affinity row had a numeric Prediction.value:

• VaccinePeptide target-epitope sorting now ignores missing/NaN IC50 or percentile-rank values and sends missing-only epitopes to the end.
• Template report rows render missing/NaN mutant IC50 as No prediction, matching the existing WT IC50 placeholder behavior.
• Test fixtures no longer derive fake percentile ranks from IC50 values.
• Bumps vaxrank.version to 3.1.8 for the PR.

Root Cause

The IC50 sort path used min(p.value for p in affinity_leaves) directly, unlike the percentile-rank path which already filtered missing values. The report path then formatted mutant IC50 with %.2f nM unconditionally.

Validation

• ./lint.sh
• ./test.sh (858 passed, 8 warnings)