more robust alert for node reports being up, while going down.

[amnonh]

When a node goes down (or in extreme situtation when a node stop to operate) it can be stack in a limbo,
it reports metric, so Prometheus think it's up, while not all metrics are reported.

We already have an alert for this situations but it uses absent that is only valid for a period of time.
Instead we shoul move to:

sum(up{job="scylla"}>0)by(instance) unless sum(scylla_transport_requests_served{shard="0"}) by(instance)