There were multiple problems:

- Some timeouts were too small. Timeout 0.1 is a very small value,
  which leads to flakiness in 100% cases sooner or later.

- One timeout was too big - 5 second waiting, whereas it could
  easily be less than a second. Anyway it was expected to fail. No
  need to wait so long.

- To check if timeout really passed whole, was used os.time(),
  which is incorrect: precision is seconds. Also the passed time
  was checked using equation duration == timeout, but it is also
  wrong. When something is blocked on a timeout, if the system is
  not real-time, the really passed time is always >= timeout. Not
  == timeout.

- In the failover test there was no fullmesh. As a result, when a
  replica was promoted and wrote something into the sync space, it
  wasn't replicated to master. But the test passed because
  1) The incorrect behaviour was in .result file;
  2) On the replica the quorum was default, i.e. 1. So the replica
   didn't wait master, and successfully wrote data into the sync
   space.

The initial problem of the test was that in the last case one of
the test jobs somehow got the old master seeing the replica's
data. But it is impossible, there was no replication from the
replica to master. Anyway now the test case is reworked, and even
if it would fail, it would be a new fail.

Closes #5168