Nodes that are down when connecting to a cluster are not considered "down"

[Andrea Leopardi]

See comments from #288 (comment) and on.

We will implement the solution described in #288 (comment):

When we start a pool of connections, we can set a timer in the control connection, like {:check_host_health, Host.t()}. After that timer expires, we can check whether we have any connections for that host in the registry. If there are no connections (which means they never registered themselves) or they're all :down, we can mark the node as down as we do now. This would overcome never getting the {:disconnected, pid} message.

@harunzengin is working on this one.