Load legacy PaperBench reproduction metadata with optional defaults

[sylvesterkaczmarek]

Summary

Keep cached PaperBench reproduction metadata loadable when it was written before newer optional fields were added.

ReproductionMetadata already declares retried_results with an empty-list default and three fields as optional, but from_dict() indexes all four keys directly. Reusing an older *_executed_metadata.json can therefore raise ValueError even though the dataclass has well-defined defaults for the missing fields.

Fixes #142.

Fix

• default missing retried_results to [];
• load repro_execution_time, git_status_after_reproduce, and executed_submission with dict.get();
• keep the core reproduction fields required as before.

Regression coverage

Adds tests verifying that:

• a legacy dictionary containing only the core required fields loads successfully with the dataclass defaults;
• newer optional fields and retried-run metadata are preserved when present.

Serialization output and the required core metadata schema are unchanged.