Fix: Avoid unnecessary producer epoch bumps

[marcin-krystianc]

Fixes: #4953

When the producer is idempotent and max.in.flight.requests.per.connection is set to a value between 2 and 5, it's normal to receive OUT_OF_ORDER_SEQUENCE_NUMBER produce responses for requests R2 through R5 when the R1 failed for any other reason.

Bumping the producer epoch in this scenario violates the "exactly-once" guarantees. Therefore, we believe that it's unnecessary to bump the producer's epoch; re-enqueuing the messages is sufficient.

The same "retry, but don't bump producer epoch" behavior is implemented in the Java client: https://github.com/apache/kafka/blob/a6a588fbed9982598377060c63f94ee6184b4295/clients/src/main/java/org/apache/kafka/clients/producer/internals/TransactionManager.java#L1015-L1016

[confluent-cla-assistant]

🎉 All Contributor License Agreements have been signed. Ready to merge.
✅ marcin-krystianc
_{Please push an empty commit if you would like to re-run the checks to verify CLA status for all contributors.}