Investigate executor broadcast shuffle handling performance on Databricks

[tgravescs]

Is your feature request related to a problem? Please describe.
Related to fix for #10165. We have a couple of specific handling for executor side broadcasts where we add in an extra shuffle to go from GpuExchange -> GpuColumnarToRow -> Exchange (EXECUTOR_BROADCAST). We do this because the BroadcastHashJoin is expecting the Exchange node there and not a ColumnarToRow exec. This works but its likely adding in extra overhead.

Perhaps we can make a special GPUExchange that has a doExecute and handles the columnar to row itself, but we should analyze the performance impact and if something like that will work.

after fix for #10165 The plan looks like:

                  +- Exchange (16)
                     +- GpuColumnarToRow (15)
                        +- GpuShuffleCoalesce (14)
                           +- ShuffleQueryStage (13), 
                              +- GpuColumnarExchange (12)
                                 +- GpuCoalesceBatches (11)
                                    +- GpuFilter (10)

You can see that we have an extra shuffle in there.