[VL] Make sure ColumnarSubqueryBroadcastExec compatible with fallen back vanilla exchange in AQE

[zhztheplayer]

This is a RAS fix but also makes the code used by legacy planner more robust too.

For input plan:

ColumnarSubqueryBroadcast dynamicpruning#1601, 0, [d_date_sk#1519L], [id=#1441]
+- AdaptiveSparkPlan isFinalPlan=false
   +- BroadcastExchange HashedRelationBroadcastMode(List(input[0, bigint, true]),false), [plan_id=1413]
      +- Project [d_date_sk#1519L]
         +- Filter ((((isnotnull(d_year#1525L) AND isnotnull(d_moy#1527L)) AND (d_year#1525L = 1999)) AND (d_moy#1527L = 1)) AND isnotnull(d_date_sk#1519L))
            +- FileScan parquet spark_catalog.default.date_dim[d_date_sk#1519L,d_year#1525L,d_moy#1527L] Batched: true, DataFilters: [isnotnull(d_year#1525L), isnotnull(d_moy#1527L), (d_year#1525L = 1999), (d_moy#1527L = 1), isnot..., Format: Parquet, Location: InMemoryFileIndex(1 paths)[file:/tmp/tpcds-generated-1000.0/date_dim], PartitionFilters: [], PushedFilters: [IsNotNull(d_year), IsNotNull(d_moy), EqualTo(d_year,1999), EqualTo(d_moy,1), IsNotNull(d_date_sk)], ReadSchema: struct<d_date_sk:bigint,d_year:bigint,d_moy:bigint>

When the project is not offloaded for some reason, RAS planner may decide to keep BroadcastExchange as a row plan to avoid transitions. In that case, ColumnarSubqueryBroadcast will fail execution.

This patch fixes the issue.

[github-actions]

Run Gluten Clickhouse CI

[github-actions]

Run Gluten Clickhouse CI