Closes #1769

Summary

• Added failover_monitor coroutine to NPUWorkerManager
• Detects dead workers via expired heartbeat TTL (npu:worker:{id}:status)
• Re-queues orphaned tasks from running→pending (with retry tracking)
• Tasks exceeding max_retries moved to failed queue
• Launches as background task alongside existing health monitoring
• All new functions under 30 lines (extracted helpers)

Test plan

• File compiles without errors
• All pre-commit hooks pass (Black, flake8, bandit)
• Uses existing Redis patterns (no new client creation)
• Respects max_retries limit
• Logs all task migrations