fix: reconcile orphaned parsing tasks after system crash/reboot

[timon0305]

Summary

Fixes #13171

When the system crashes or reboots during document parsing, Redis loses queued tasks while MySQL retains the "Parsing" state (run='1', progress > 0). This causes documents to appear stuck in "Parsing" indefinitely with no backend processing.

This PR adds startup reconciliation logic to detect and re-queue orphaned tasks:

• DocumentService.get_orphaned_parsing_docs() - Queries documents stuck in Parsing state (run=RUNNING, 0 < progress < 1)
• TaskService.get_incomplete_tasks_by_doc_ids() - Retrieves incomplete tasks for orphaned documents
• requeue_orphaned_tasks() - Reconciliation function that re-queues incomplete tasks into Redis, or marks documents as failed if no tasks remain
• Startup hook in task_executor.py - Uses a Redis distributed lock so only one worker runs reconciliation, preventing duplicate re-queuing

Documents with no remaining incomplete tasks are marked as FAIL with an explanatory message rather than left in a permanently stuck state.

Test plan

• Deploy with documents in parsing state, restart services, verify orphaned documents resume parsing
• Verify only one worker runs reconciliation via distributed lock
• Verify documents with no incomplete tasks are marked as FAIL

[dosubot]

Related Documentation

Checked 0 published document(s) in 1 knowledge base(s). No updates required.

^{How did I do? Any feedback?}