(WIP)KAFKA-9965/KAFKA-13303: RoundRobinPartitioner broken by KIP-480

[jim0987795064]

RoundRobinPartitioner behaviour was broken by sticky partitioning (KIP-480).

This patch addresses the behavioural issue caused by the second call to partition() after onNewBatch(), in a predicatable and thread-safe manner.

Unit tested by simulation of multiple threads producing to two topics with race conditions.

Changes:

• Store the last used partition per batch in a thread-safe queue to ensures thread-safety across multiple producer threads.
• Adds unit tests that simulate multiple threads producing to two topics.

Reasons:

• The original round-robin algorithm relies on a shared counter, which can result in multiple threads assigning the same partition for new batches. This breaks the intended even distribution and causes partition skew.

[ash-at-github]

It's great that this test verifies the next partition selected matches the enqueued value for each topic. Consider adding test cases for some edge cases, like empty queue and error handling