…fter rescaling

Purpose of change

ActionExecutionOperator persists currentProcessingKeysOpState as union operator state so all subtasks can see in-flight keys after restore/rescaling. However, tryResumeProcessActionTasks() resumed every restored key without checking whether the key belonged to the current subtask.

That behavior is incorrect after scale-out: multiple subtasks can receive the same union state entry, so non-owner subtasks may also enqueue resume work for keys they do not own. This can lead to duplicated processing after recovery.

This change fixes the restore path by filtering currentProcessingKeysOpState with subtask ownership before resubmitting work:

• compute the key group for each restored key
• compute the owning subtask for the current maxParallelism and parallelism
• only resume keys owned by the current subtask

The ownership check is encapsulated in isKeyOwnedByCurrentSubtask(...) in ActionExecutionOperator.java.

Tests

Added an operator-level restore/rescaling regression test in ActionExecutionOperatorTest.java:

• snapshot state from parallelism=1
• repartition it to parallelism=2
• restore both subtasks
• verify only the owner subtask gets a resume mail, while the non-owner subtask does not

Documentation

• doc-needed
• doc-not-needed
• doc-included