Publish a FIBER optical alert to NVSentinel as a HealthEvent

Body: {"alert_id": 123} or {"switch": "leaf-010", "port": "Ethernet1/41"} (port "-" = switch-level telemetry_silence). The alert is mapped to one datamodels.HealthEvent per GPU node whose uplink is that port (gpu_nodes.port_id). Dry-run (payload returned, nothing sent) when FIBER_NVSENTINEL_PUBLISH_URL is unset or dry_run: true. Real publishes are deduplicated per (node, check) for 1 h unless force: true; a failed send releases the dedup budget so the retry is not suppressed. Optional hours (default 24) bounds the alert lookup; optional message/alert_type/severity synthesise an alert when none is on record — that synthetic path is dry-run/demo only: it is forced to dry-run (unless FIBER_NVSENTINEL_ALLOW_SYNTHETIC_PUBLISH=1) and never sets isFatal, so a fabricated event cannot drive an NVSentinel cordon/drain.