revert(worker): rely on visibility-timeout recovery

[suguanYang]

Summary

• Remove the bounded gevent active-task cancellation introduced by fix(worker): bound gevent warm shutdown #282.
• Retain the reconnect-safe gevent cancellation no-op and document why interruption recovery relies on Redis visibility recovery.
• Retain the independent watchdog, the 4,500-second visibility contract, and the existing billing/idempotency behavior.

Staging evidence

During the approved active-job drill, job job_07feda62dfc9 was processing when ECS sent SIGTERM. The 90-second callback cancelled the greenlet and removed its workspace while a PyMuPDF subprocess was still writing. The process still missed the ECS stop deadline, exited 137, and did not restore the reservation promptly. The owner accepted the already-tested visibility-timeout fallback, with up to approximately 75 minutes of interruption delay, as the migration reliability trade-off.

Validation

• 14 passed across worker shutdown, visibility recovery, watchdog, bootstrap, and health contracts
• Ruff passed for changed Python files
• Pyright passed for worker_bootstrap.py
• Standards/spec review found no blocking findings; unrelated bootstrap extraction remains out of scope

Tracks Ontos-AI/knowhere-api-infra#22.