Bug 893851

This is causing regression tests to fail most of the time at step 12 of bug-879490.t and is also readily reproducible in isolation.  The problem is that we're adding the new peer to our peer list *after* calling glusterd_friend_rpc_create.  Unfortunately, by that time the disconnect-notify callback associated with our new  RPC object might already have been invoked, removing *and freeing* the new peer.  Thus, we add it after freeing it, leading to a corrupt list and a likely future crash - most often in glusterd_op_ac_send_lock while deleting the volume.  The test fails because glusterd disconnects and dies in mid-operation.

The fix is to change the order of addition to and deletion from the peer list.  In the process, the general control flow of glusterd_friend_add was simplified so that it's not only correct but more readily verified to be so.