Handle uneven PaperBench seed counts when parsing runs

[sylvesterkaczmarek]

Summary

Allow parse_run_data() to build incomplete PaperBench evaluation runs when papers have uneven numbers of available seeds instead of indexing past shorter histories.

The parser computes the maximum seed count across papers, but its short-paper guard currently skips only when seed == len(data). If one paper has one run and another has three, seed 2 satisfies 2 != 1 and the code attempts data[2], raising IndexError.

Fixes #140.

Fix

Skip whenever seed >= len(data).

Regression coverage

Adds a unit regression with one paper containing one run and another containing three runs. It verifies parsing produces three EvaluationRun objects, with the shorter paper present only in the first run and no out-of-range access.

Timestamp ordering, recent-seed selection, disqualification handling, and complete-run filtering are unchanged.

[sylvesterkaczmarek]

Superseded by #150. The run-group-based parser in #150 removes the positional per-paper indexing entirely, so uneven paper histories remain incomplete within their actual run group instead of causing out-of-range access or borrowing records from another run. #150 now explicitly fixes #140 as well.