There was a test case about master writing a transaction, replica
stating snapshot, master writing rollback, and replica canceling
the snapshot. Because its data was spoiled with rolled back data.

Sometimes it could happen that the master started writing the
transaction to WAL in a newly created fiber, then the test
switched to replica, successfully created the snapshot, and only
then the data from master was received. As a result, the following
rollback didn't affect the already finished snapshot.

The patch forces the replica wait for receipt of dirty data from
master before starting the snapshot.

Closes #5167