Previously, we'd keep retrying a connection until the channel closed.
With this change, we only retry connecting *if there's a watcher on the
subchannel connection state*. This ultimately means that if the
lb_policy doesn't care if the subchannel connects, it'll stop trying.