test_replay: Fix hang since confluent-kafka 1.6 release

Side effect of the following commit in librdkafka 1.6:
https://github.com/edenhill/librdkafka/commit/f418e0f721518d71ff533759698b647cb2e89b80

Tests was relying on a buggy behavior of the mocked kafka cluster: two
subsequent consumers setup with the same group id should receive a
different set of messages, rather than the same set of messages.

Also explicitly commit messages once consumed.