Spark: Faster net changelogs using identifier columns

[1raghavmahajan]

Closes #14249

Implementation

The implementation is similar to what was initially proposed:

1. Repartition by identifier_columns and sort within partition by identifier_columns + change_ordinal
2. Apply RemoveCarryoverIterator.

Note: Above is the same as net_changes without identifier columns but with a simpler repartition spec.

3. Use window functions to identify first and last changes for each logical row
4. Filter to keep only first and last changes (as per change_ordinal) for each logical row

Note: Above performs the netting of the changes, we get rid of all change except for the first and last change ordinal, this is cheaper than iterating through them all. Existing net_changes cannot leverage this as we do not have a consistent set of identifier columns across the entire snapshot range so we need to iterate through them all to build the lineage.

5. Remove INSERT-DELETE (no-op) pairs using an iterator.
6. Calculate pre/post images using first DELETE - last INSERT pairs.

Note: Above is similar to existing ComputeUpdateIterator). Here we need to handle multiple INSERTS/DELETEs entries(as the intermediate changes aren't present).

Testing

• Added iterator tests for ComputeNetUpdateIterator and RemoveNoopPairIterator
• Updated integration tests for CreateChangelogViewProcedure

[1raghavmahajan]

cc @flyrain @szehon-ho @aokolnychyi @RussellSpitzer