Since due to SERVER-45610, we can successfully execute some reads of nodes in REMOVED, ROLLBACK, and RECOVERING states, to get a consistent initial sync we must ensure we never connect to such nodes in the first place.

This ticket also ensures we do correct backoff when re-connecting (whether due to a node in the wrong state, or other reasons). And as a side effect it eliminates the ambiguity of the NotMasterOrSecondary error (which may be permanent due to a removed node, or temporary due to a restart); since we will never connect to a node a state other than PRIMARY or SECONDARY, we will never see that error.

The testing for this is in SERVER-42705 and SERVER-45423.