Add Daft audio stability benchmark for Ray

[FANNG1]

Summary

• add a standalone benchmark CLI for data generation, Mock LLM serving, direct Ray execution, Ray Jobs submission, and report generation
• exercise the full MinIO audio download, SenseVoice ASR, Mock LLM, and Lance Blob v2 path with per-row failure status
• collect Daft event logs and per-node resource samples, apply memory SLO reporting, partition manifests for distributed execution, and preserve reports on a configurable shared path
• document local usage, one-hour sizing, and scaling to a 4-worker Ray cluster without starting Docker or MinIO

Validation

• uv run pytest -q: 114 passed, 1 deselected
• real local-smoke --count 4 against the configured MinIO and local Ray: 4/4 rows succeeded and Blob v2 validation passed
• smoke processed 297.39 audio seconds in 146.39 wall seconds; ASR P95 was 4.55s and Ray process RSS peaked at about 5.11GB

Notes

• the real end-to-end run has been validated locally; the documented 4-worker cluster configuration has not yet been exercised on a live multi-node cluster
• benchmark data and Lance outputs use an isolated s3://benchmark/audio/<run-id>/ prefix