Get rid of the JOIN_PRIMARY state. (#796)

* Get rid of the JOIN_PRIMARY state.

We still need to maintain backward compatibility of course, in case there is
an upgrade where one node is in JOIN_PRIMARY and then we want to be able to
continue working with the node.

Now, any node that reports JOIN_PRIMARY is unconditionnally assigned PRIMARY
instead, and we never assign JOIN_PRIMARY anymore. We still have some code
that matches the state in the group level FSM, and we keep the full compat
on the client side for the moment.

* Per review, preserve our WAIT_PRIMARY to PRIMARY transition.

* Per review, APPLY_SETTINGS is only required when replicationQuorum.

docs/failover-state-machine.rst

@@ -67,8 +67,7 @@ all the other nodes.
 
 In the previous diagram we can see that we have a list of six states where
 the application can connect to a read-write Postgres service: ``single``,
-``wait_primary``, ``primary``, ``prepare_maintenance``, ``apply_settings``,
-and ``join_primary``.
+``wait_primary``, ``primary``, ``prepare_maintenance``, and ``apply_settings``.
 
 Init
 ^^^^
@@ -111,6 +110,12 @@ allows the primary node to apply necessary changes to its HBA setup before
 allowing the new node joining the system to run the ``pg_basebackup``
 command.
 
+.. important::
+
+   This state has been deprecated, and is no longer assigned to nodes. Any
+   time we would have used ``join_primary`` before, we now use ``primary``
+   instead.
+
 Primary
 ^^^^^^^