[FEA] Support SubqueryBroadcast on GPU to enable exchange reuse during DPP

[sperlingxx]

Is your feature request related to a problem? Please describe.
In current, when GPU overrides are applied, DPP can not reuse BroadcastExchange as subquery. The reason is that rapids plugin will fall back the broadcast exchange for subquery into CPU, since we have not supported SubqueryBroadcastExec on GPU yet. Meanwhile, the original broadcast exchange, which is supposed to be reused for subquery, is converted into GpuBroadcastExchange. Therefore, the planned reuse of exchange doesn't carry on in terms of the final plan.

Here is a comparison between CPU plan and GPU plan of DPP through broadcast reuse:

*(3) HashAggregate(keys=[key#156], functions=[sum(value#155)], output=[key#156, sum(value)#161L])
+- Exchange hashpartitioning(key#156, 12), ENSURE_REQUIREMENTS, [id=#221]
   +- *(2) HashAggregate(keys=[key#156], functions=[partial_sum(value#155)], output=[key#156, sum#166L])
      +- *(2) Project [value#155, key#156]
         +- *(2) BroadcastHashJoin [key#156], [key#157], Inner, BuildRight, false
            :- *(2) Filter (isnotnull(value#155) AND (value#155 > 0))
            :  +- *(2) ColumnarToRow
            :     +- FileScan parquet default.tmp_table_574386_0[value#155,key#156] Batched: true, DataFilters: [isnotnull(value#155), (value#155 > 0)], Format: Parquet, Location: InMemoryFileIndex(10 paths)[file:/home/alfred/workspace/codes/spark-rapids/integration_tests/targ..., PartitionFilters: [isnotnull(key#156), dynamicpruningexpression(key#156 IN dynamicpruning#164)], PushedFilters: [IsNotNull(value), GreaterThan(value,0)], ReadSchema: struct<value:int>
            :           +- SubqueryBroadcast dynamicpruning#164, 0, [key#157], [id=#126]
            :              +- BroadcastExchange HashedRelationBroadcastMode(List(cast(input[0, int, true] as bigint)),false), [id=#125]
            :                 +- *(1) Project [key#157]
            :                    +- *(1) Filter ((isnotnull(filter#159) AND (filter#159 = 1552)) AND isnotnull(key#157))
            :                       +- *(1) ColumnarToRow
            :                          +- FileScan parquet default.tmp_table_574386_1[key#157,filter#159] Batched: true, DataFilters: [isnotnull(filter#159), (filter#159 = 1552), isnotnull(key#157)], Format: Parquet, Location: InMemoryFileIndex(1 paths)[file:/home/alfred/workspace/codes/spark-rapids/integration_tests/targe..., PartitionFilters: [], PushedFilters: [IsNotNull(filter), EqualTo(filter,1552), IsNotNull(key)], ReadSchema: struct<key:int,filter:int>
            +- ReusedExchange [key#157], BroadcastExchange HashedRelationBroadcastMode(List(cast(input[0, int, true] as bigint)),false), [id=#125]

GpuColumnarToRow false
+- GpuHashAggregate(keys=[key#156], functions=[gpusum(value#155, LongType)], output=[key#156, sum(value)#161L])
   +- GpuShuffleCoalesce 2147483647
      +- GpuColumnarExchange gpuhashpartitioning(key#156, 12), ENSURE_REQUIREMENTS, [id=#199]
         +- GpuHashAggregate(keys=[key#156], functions=[partial_gpusum(value#155, LongType)], output=[key#156, sum#166L])
            +- GpuProject [value#155, key#156]
               +- GpuBroadcastHashJoin [key#156], [key#157], Inner, GpuBuildRight
                  :- GpuCoalesceBatches targetsize(2147483647)
                  :  +- GpuFilter (gpuisnotnull(value#155) AND (value#155 > 0)), true
                  :     +- GpuFileGpuScan parquet default.tmp_table_625824_0[value#155,key#156] Batched: true, DataFilters: [isnotnull(value#155), (value#155 > 0)], Format: Parquet, Location: InMemoryFileIndex[file:/home/alfred/workspace/codes/spark-rapids/integration_tests/target/run_dir..., PartitionFilters: [isnotnull(key#156), dynamicpruningexpression(key#156 IN dynamicpruning#164)], PushedFilters: [IsNotNull(value), GreaterThan(value,0)], ReadSchema: struct<value:int>
                  :           +- SubqueryBroadcast dynamicpruning#164, 0, [key#157], [id=#136]
                  :              +- BroadcastExchange HashedRelationBroadcastMode(List(cast(input[0, int, true] as bigint)),false), [id=#135]
                  :                 +- GpuColumnarToRow false
                  :                    +- GpuProject [key#157]
                  :                       +- GpuCoalesceBatches targetsize(2147483647)
                  :                          +- GpuFilter ((gpuisnotnull(filter#159) AND (filter#159 = 1552)) AND gpuisnotnull(key#157)), true
                  :                             +- GpuFileGpuScan parquet default.tmp_table_625824_1[key#157,filter#159] Batched: true, DataFilters: [isnotnull(filter#159), (filter#159 = 1552), isnotnull(key#157)], Format: Parquet, Location: InMemoryFileIndex[file:/home/alfred/workspace/codes/spark-rapids/integration_tests/target/run_dir..., PartitionFilters: [], PushedFilters: [IsNotNull(filter), EqualTo(filter,1552), IsNotNull(key)], ReadSchema: struct<key:int,filter:int>
                  +- GpuBroadcastExchange HashedRelationBroadcastMode(List(cast(input[0, int, true] as bigint)),false), [id=#195]
                     +- GpuProject [key#157]
                        +- GpuCoalesceBatches targetsize(2147483647)
                           +- GpuFilter ((gpuisnotnull(filter#159) AND (filter#159 = 1552)) AND gpuisnotnull(key#157)), true
                              +- GpuFileGpuScan parquet default.tmp_table_625824_1[key#157,filter#159] Batched: true, DataFilters: [isnotnull(filter#159), (filter#159 = 1552), isnotnull(key#157)], Format: Parquet, Location: InMemoryFileIndex[file:/home/alfred/workspace/codes/spark-rapids/integration_tests/target/run_dir..., PartitionFilters: [], PushedFilters: [IsNotNull(filter), EqualTo(filter,1552), IsNotNull(key)], ReadSchema: struct<key:int,filter:int>

Describe the solution you'd like
Implement GpuSubqueryBroadcastExec, which is able to reuse GpuBroadcastExchangeExec in row-wise view (Array of UnsafeRow).