Optimize creating indexes when writing partitioned table

[malhotrashivam]

In the current code when writing a source table in a key-value partitioned fashion, if user provides indexing columns, we compute the constituent tables to write and then add an index on each of the provided column for each constituent table.

This can be optimized by intersecting with any existing indexes on the original table. We can perform a transform to intersect the index tables' row sets with the constituent row sets, then filter the empty row sets.

Pseudocode for intersecting (courtesy of @rcaudy), probably with bad parentheses, etc:

partitionedSource = sourceTable.partitionBy(partitioningColumns)
for dataIndex in sourceTable indexes:
    indexTable = dataIndex.table()
    partitionedIndexTable = partitionedSource.transform(c ->
        // Do this with `FunctionalColumns` to avoid compiles
        indexTable
            .update(List.of(new FunctionalColumn(dataIndex.rowSetColumnName(), RowSet.class, dataIndex.rowSetColumnName(), RowSet.class, dataIndexRowSet -> dataIndexRowSet.intersect(c.getRowSet()))))
            .updateView(List.of(new FunctionalColumn(dataIndex.rowSetColumnName(), RowSet.class, "__NON_EMPTY__", Boolean.class, RowSet::isNonempty)))
            .where(Filter.isTrue(ColumnName.of("__NON_EMPTY__"))))
    // We *could* add `StandaloneDataIndex`es to the constituents, or just write from the partitioned index table

(Found during #5105)