Reset restart_count for host-network workers stuck in Creating with a live container

[Shine-neko]

Problem

A worker in network: host mode can never satisfy Ring's readiness checks (its container has no resolvable address), so it stays Creating for its entire life even though its container runs fine.

The anti-flap window that refills restart_count only reset the count when the deployment reached Running. A host-network worker therefore accumulated a monotonic restart_count: its retry backoff climbed to the 60s cap and never came back down. Every redeploy got progressively slower (observed: ~5 min to recreate a container after a drop).

Fix

Treat a deployment as healthy-running for the anti-flap window whenever it actually has a live container, not only when it reached Running: Running, or Creating with a non-empty instance list. Extracted into a small pure has_live_container() helper.

Impact

Measured in production after deploying this fix: container recreation after a drop dropped from ~5 min to ~20s.

Tests

Added 4 unit tests covering has_live_container (Running, Creating+instance, Creating without instance, other states). Full suite: 735 passed.