Add support for parallel scoring with --num_workers option

[RogerTzeng]

• Introduced --num_workers argument in scorer.py for CPU-based parallel processing.
• Implemented parallel scoring logic in scorer_shared.py to enhance performance.
• Added tests to validate functionality of parallel scoring and ensure input order preservation.

[ftshijt]

Thanks for your contribution! There are some issues I would like to point:

• versa/scorer_shared.py:554: the parallel resume path rewrites the whole JSONL file with _write_jsonl_scores(..., "w").

  This changes --resume from append-safe to rewrite-at-end. If the process is interrupted during the final write, or the write fails, previously completed resume results can be lost or truncated. It also means newly completed parallel work is not checkpointed until all jobs finish.

  I’d suggest either appending safely like the serial path, or writing to a temp file and atomically replacing the output only after a successful full write.

• --num_workers > 1 with --scoring_mode metric or corpus-only configs is accepted but ignored.

  A warning or parser error would make that clearer, since the README says those modes remain serial.

After the above issues are addressed, please also make sure you pass the CI.