feat(scheduler): RayLogCleanupTask 4-stage cleanup of session_latest/logs

[jinbai340997]

Summary

Modified files

• rock/admin/scheduler/tasks/ray_log_cleanup_task.py — 4-stage shell pipeline + daemon whitelist + 3 new tunable params
  (live_log_keep_days, old_logs_keep_hours, setup_log_keep_minutes) + per-category counters in return dict
• tests/unit/admin/scheduler/test_ray_log_cleanup_task.py — 29 tests covering all 4 stages, daemon whitelist regression guards
  (especially PART 2a), output parsing, init validation, from_config defaults

New tunables (all backward compatible, defaults preserve existing behavior)

• live_log_keep_days — PART 2b mtime threshold for non-PID non-daemon files (default 7)
• old_logs_keep_hours — PART 3 mtime threshold for logs/old/ files (default 24)
• setup_log_keep_minutes — PART 2c mtime threshold for runtime_env_setup-* (default 60)

Cron mirror (out of scope of this PR)

The ray-head node runs an identical shell pipeline via cron daily (no rocklet on ray-head, scheduler doesn't reach it). The shell
mirror lives in the rock-internal repo and is kept lock-step with this task's pipeline (manual sync — both stages added together).

Test plan

• Unit: uv run pytest tests/unit/admin/scheduler/test_ray_log_cleanup_task.py -v — 29/29 PASS
• Pre-prod worker 11.8.74.237 (6-month Ray session, 1705 files):
  • 160 daemon files preserved (BEFORE=160 → AFTER=160, zero diff) ← validates fix
  • 41 dead-PID worker files removed in single run (PART 2a)
  • runtime_env_setup-<hex> files correctly handled by new PART 2c (verified 770+ such files exist on production worker)
  • PART 3 verified by synthetic test: touched 25h-old file → deleted, fresh file → preserved
• Regression: ran the task on a long-lived session that had agent-1183572637.err etc. — agent / log_monitor / dashboard_agent
  files preserved as expected
  fixes [Feature] RayLogCleanupTask 4-stage cleanup of session_latest/logs (PID-aware + daemon-safe) #1028