[codex] Combine separate predictor results

[iskandr]

Closes #170.

Summary

• Add combine_predictor_results() for stacking separate long-form predictor outputs into the same shape as a multi-model TopiaryPredictor run.
• Preserve model-version provenance from fresh TopiaryPredictor DataFrame outputs via lightweight DataFrame.attrs metadata.
• Validate matching (peptide, allele) identity sets and reject duplicate prediction_method_name values across inputs.
• Keep emitted rows as the source of truth for which predictor produced which quantities: prediction_method_name, predictor_version, kind, and value/rank columns are not duplicated into kind_support metadata.
• Preserve structured Metadata.extra values through TSV/CSV comment-block round trips using explicit JSON markers, with legacy kind_support parsing only for existing files that already wrote Python-dict-looking metadata.
• Document the separate-run combine workflow and bump version to 5.16.2.

Validation

• pytest tests/test_combine_predictor_results.py tests/test_io.py::TestParseCommentBlock tests/test_io.py::TestFormatCommentBlock
• pytest tests/test_combine_predictor_results.py tests/test_result.py tests/test_io.py tests/test_io_pvacseq.py tests/test_kind_support.py tests/test_wide.py tests/test_best_allele.py
• ./lint.sh
• git diff --check
• ./test.sh

[coveralls]

coverage: 89.828% (+0.001%) from 89.827% — combine-predictor-results into master