ReplicationSink should create Connection instances lazily

Over on HBASE-12911 I have a patch that registers Connection instances with the metrics system. In both standalone server and tll client applications, I was surprised to see multiple connection objects showing up that are unused. These are pretty heavy objects, including lots of client threads for the batch pool. We should track these down and remove them – if they're not some kind of phantom artifacts of my WIP patch over there.