worker: make drain deadlines explicit

[PaulFidika]

Summary

• make Drain(0) truly non-expiring while results are in flight or waiting to reconnect
• treat a positive deadline as one total monotonic budget beginning when Drain is received
• stop admission synchronously before a back-to-back RunJob can enter
• close abruptly, rather than claiming a clean flush, when a positive deadline expires with results pending
• preserve authentication, protocol, HelloAck, and ordinary non-drain disconnect deadlines

Validation

• 22 focused lifecycle and real gRPC tests passed
• Ruff passed for all changed Python files
• Mypy passed for changed source and the new test module
• HTTP-timeout structural lint passed
• git diff check passed

The focused test run used the existing uv-managed development environment with --no-sync. A fresh frozen sync is currently blocked by the pre-existing torch 2.13.0+cu130 lock hash mismatch; this PR does not change dependencies or the lockfile.