Health-check load balancer backends and drop unhealthy ones from the Maglev table #704

Summary

datum-cloud/enhancements#884 requires the load balancer to spread traffic across active, healthy endpoints. Today a NetworkRule's backend list is static. When a backend dies, it keeps its slot in the Maglev table and every flow hashed to it fails until someone edits the rule.

The constraint: every gateway node must build the same Maglev table, or a flow that moves between nodes lands on a different backend. Health can't be decided locally on each node. The healthy set has to be shared state that every node reads.

Acceptance criteria

• A short design doc covers: who probes (each gateway node, one per region, or the backend's own node), what the probe is (TCP connect, UDP), where the result lives (rule status or a new CR), and how flapping is damped
• Unhealthy backends are removed from the input to the Maglev table on every gateway node, and restored when they recover
• Per-backend health appears in status for Public API for network load balancer #696 to expose
• A lab test kills a backend and confirms new flows avoid it without operator action

Phase 2 of #698.