[BUG] NDS query 64 falls back to CPU only for a shuffle

[jlowe]

When running NDS query 64 at the 3k scale factor on AWS EMR 6.12, there is an unnecessary fallback to the CPU for a shuffle. The node sequence in the plan UI is:

GpuHashAggregate(partial) ->
  GpuColumnarToRow ->
    AdaptiveSparkPlan ->
      Exchange ->
        ReusedExchange ->
          AQEShuffleRead ->
            GpuRowToColumnar ->
              GpuHashAggregate(final)

[sameerz]

Need to check for this in an on-premises cluster. There were no notifications in the driver log.

Marking as high priority for the interim while we investigate.

[jlowe]

This is also happening in NDS query 23_part1, 23_part2, 24_part1, 24_part2, 47, and 57.