Fix bench-api deploy on single-node Swarm (stop-first rollout)

[Navneethd8]

Summary

• Use stop-first rolling updates for bench-api and bench-worker on single-node Swarm (EC2 prod) instead of start-first, which requires 2× memory reservations and triggers insufficient resources
• Reconcile extra running tasks before deploy (cleans up failed start-first rollouts that left multiple bench-api containers)
• Tighten rollout completion check: require exact replica count, not >=

Test plan

• On prod EC2: ./s/ops/deploy.sh bench-api succeeds after merge
• docker service ps bench-api shows exactly 1 running task after deploy
• docker ps lists only one bench-api.* container

Immediate prod recovery (before merge)

docker service update --force --update-order stop-first bench-api
docker service ps bench-api

Made with Cursor