net.replicaset: fix a flacky test

The test checks that net.replicaset connection switches to the
new leader when the new leader is elected. Unfortunately the
test expects that after successful call of box.ctl.promote()
the old leader is guaranteed to be in RO state. Actually this
is incorrect: the new leader can legally rely on the quorum
with the third instance without even notifying the original
leader about new election term started, so the old leader may
remain in RW state. Since the replicaset connection updates
states of instances asynchronously it also may think (for some
time) that the old leader is the only RW instance in replicaset.

Fix it by including check of the new leader into retrying block
of the test.

Hotfix #9823

NO_CHANGELOG=fix flaky test
NO_DOC=fix flaky test