Keep workers reconnecting across hub outages

[PaulFidika]

What changed

• remove the 600-second disconnected worker self-exit and its WORKER_DISCONNECTED_TIMEOUT_S configuration
• keep the existing jittered reconnect loop active until explicit Drain or process stop
• delete the drain-only timeout bypass that became dead code
• add a real gRPC regression that stops the hub, observes repeated reconnect attempts, rebinds the same port, and verifies a fresh Hello without worker exit

Root cause

The worker owned a fixed disconnected lifetime even though the protocol contract requires reconnecting after every non-Drain stream loss. Once that lifetime elapsed, PID 1 exited while the provider pod could remain allocated and appear RUNNING but idle. The live four-minute incident did not reach the old 600-second default, so this removes a proven lifecycle defect without claiming it was the only cause of that event.

Validation

• pytest tests/test_drain_flush.py tests/test_token_refresh.py tests/test_worker_grpc_e2e.py: 24 passed
• mypy src: clean across 114 files
• ruff check: clean
• full suite: 1,078 passed, 13 skipped, one unrelated CUDA residency assertion; the same assertion fails on unchanged master in this host environment
• git diff --check: clean