Ensure SentencePiece tokenizer returns string pieces

[RNA4219]

Summary

• tighten the SentencePiece processor test double to require out_type=str and capture encode_as_pieces usage
• add regression coverage for both direct encode and encode_as_pieces fallback paths
• update the CLI SentencePiece tokenizer to request string pieces via encode(..., out_type=str) or fall back to encode_as_pieces

Testing

• pytest tests/quality/evaluator/test_cli.py

---

https://chatgpt.com/codex/tasks/task_e_6907d6022a7c8321a567341f58b3606c