GH-22: parallel health checks

[blainemotsinger]

Summary

Health-check rounds now run in parallel across nodes, so a slow or hung node can no longer stall health-state convergence of the rest of the fleet. Cycle time becomes the slowest node instead of the sum of all nodes.

• Parallel fan-out — checkAll spawns one goroutine per node and blocks on a sync.WaitGroup round barrier; rounds never overlap (effective interval = max(health_interval, round_duration)).
• Per-node admission guard — skip-if-busy tryStart/finish serializes same-node checks, closing the known Recheck/cycle double-count race on ConsecutiveFailures.
• Panic containment — a recovered runCheck wrapper (log + continue) means one node's panicking probe can neither crash the gateway nor wedge the guard/barrier.
• Shutdown — Stop() continues to drain the in-flight round (bounded by the 5s client probe timeout).

No config changes, no public API changes; probes (vLLM /health + gpu-exporter /metrics) stay sequential within a node. Accepted residual: a Recheck finishing just before a tick can produce at most two sequential probes for one node per interval (each count is a real failed probe).

Testing

• 5 new tests using real httptest backends: no-stall with a 2s slow node, admission guard blocks concurrent checks, Recheck no-op while in flight, Stop drains in-flight round, panic releases guard and round.
• go test ./... -race -count=1 all green; timing-sensitive tests stable at -count=5; go vet clean.

Review

• Implemented task-by-task with per-task spec + quality reviews (all clean), plus a final whole-branch review: 0 Critical / 0 Important.
• Known follow-up candidates (not gating): one-line non-blocking stop-check on the ticker branch to avoid one extra round after Stop; a panic-count metric for the recovered-panic path.