Support resumable scoring

[ftshijt]

Summary

Adds resume support for utterance-level scoring so interrupted runs can continue from an existing JSONL output file instead of starting over.

Details

• Adds a --resume flag to the standard and chunked scorer entrypoints.
• Loads completed utterance keys from an existing output file and skips them during scoring.
• Appends new results safely, including when the existing file is missing a trailing newline.
• Keeps resumed rows in the returned score list so summary computation includes previous and new results.
• Adds MetricSuite.__len__ so legacy chunked scoring code can safely check loaded metric counts.
• Adds a focused regression test for skipping completed keys.

Validation

• conda run -n versa-dev python -m pytest test/test_pipeline/test_base_metrics_pipeline.py::test_score_utterances_resume_skips_completed_keys -q
• conda run -n versa-dev python -m pytest test/test_pipeline/test_base_metrics_pipeline.py::test_score_utterances_resume_skips_completed_keys test/test_pipeline/test_base_metrics_pipeline.py::test_stoi_and_signal_pipeline_with_registry -q
• conda run -n versa-dev python -m versa.bin.scorer --pred test/test_samples/test2.scp --gt test/test_samples/test1.scp --score_config egs/separate_metrics/snr_related.yaml --output_file /private/tmp/versa-resume-cli.jsonl --io soundfile --resume twice; output stayed at 1 row
• conda run -n versa-dev python -m versa.bin.scorer_chunk --pred test/test_samples/test2.scp --gt test/test_samples/test1.scp --score_config egs/separate_metrics/snr_related.yaml --output_file /private/tmp/versa-resume-chunk.jsonl --io soundfile --enable_chunking --chunk_duration 0.5 --hop_duration 0.5 --resume twice; output stayed at 2 rows
• python3 -m py_compile versa/definition.py versa/scorer_shared.py versa/bin/scorer.py versa/bin/scorer_chunk.py test/test_pipeline/test_base_metrics_pipeline.py
• git diff --check