atenet: two replicas + PDBs for router and dns

[NekoPunch (orangeCatDeveloper)]

Summary

The atenet-router and dns Deployments are single-replica SPOFs: every deploy, node drain, or eviction takes the whole data plane down (measured ~5.5 s of hard connection failures per router pod kill; 6,880 dial errors during one GKE node-pool event — see #645). This bumps both to 2 replicas with a surge-only rollout (maxUnavailable: 0), an Envoy readiness probe (admin /ready, which passes only after the initial xDS load from the router container), preStop drain sleeps on all three containers, soft hostname spread, and PodDisruptionBudgets.

preStop uses the native K8s ≥1.30 sleep action because the ko and coredns images are distroless (no sleep binary). The router container needs its own drain: it hosts ext_proc, and if it dies while the draining Envoy still receives traffic, requests fail with 500.

The multi-replica xDS question flagged in #645 is resolved by #699 (restart- and replica-unique version epochs); in standalone sidecar mode each Envoy only talks to its pod-local router, so replicas never share an xDS stream.

Out of scope: the managed-Envoy path (envoyrunner.go) stays single-replica — #635 removes that mode entirely.

Fixes #645

Test plan

On a kind cluster, with a 5 req/s in-cluster probe against a running actor:

• rolling restart under traffic: 278/278 requests OK (single-replica baseline: ~5.5 s outage)
• force-delete one router replica under traffic: 183/183 OK
• dns rolling restart: actor domain still resolves to the router Service; no FailedPreStopHook events
• kustomize strict render of both manifests passes