Improve task manager reliability

[renkun-ken]

Summary

• isolate task callback and polling failures so they cannot terminate the server or leak scheduler capacity
• provision persistent workers according to queued demand and dispatch recent tasks newest-first
• make task-manager shutdown best-effort and idempotent
• add regression coverage for failure handling, scaling, recency, and shutdown

Tests

• task tests: 50 passed
• full suite: 2,078 passed, 1 skipped