Deduplicate message pieces before batch scoring

[biefan]

Summary

• apply BatchScorer._remove_duplicates() before grouping filtered message pieces into conversations
• fail fast when filtering leaves no original message pieces to score
• add a regression test that confirms duplicate message pieces are not forwarded to batch scoring

Problem

BatchScorer.score_responses_by_filters_async() fetches message pieces from memory and groups them into conversations immediately, even though the class already has a _remove_duplicates() helper. When both original and duplicate MessagePiece rows are returned, duplicate pieces are forwarded to score_prompts_batch_async() and can be scored multiple times.

Validation

• .venv/bin/pytest tests/unit/score -q
• .venv/bin/ruff check pyrit/score/batch_scorer.py tests/unit/score/test_batch_scorer.py