Prepare deployable cluster shape for gpu_control_plane e2e

[skel84]

Context

The GPU Control Plane repo now has a packaging and release path, but a real-cluster e2e test still needs a deployable AllocDB service.

Related docs:

• docs/real-cluster-e2e-roadmap.md
• docs/operator-runbook.md

Roadmap

• package a deployable service shape for the replicated node
• make persistence layout explicit for WAL, snapshots, and replica metadata
• expose cluster health clearly through readiness/liveness and metrics
• document startup, restart, isolate/heal, failover, and rejoin flows
• prove a minimal in-cluster smoke for submit/read and restart/rejoin safety

Acceptance

• AllocDB can be deployed from documented manifests or overlays
• the deployed service survives restart with durable state intact
• a minimal smoke passes against the deployed service
• the operational runbook matches the deployed shape

[skel84]

Paired roadmap item: skel84/gpu_control_plane#35.