[BUG] GpuShuffleExchangeExec in v2 shims has inconsistent packaging

[jlowe]

GpuShuffleExchangeExec was moved into the v2 shim style in #3559, but this class appears in different packages depending upon which shim is being used:

$ find . -name GpuShuffleExchangeExec.scala | xargs grep ^package
./sql-plugin/src/main/311+-nondb/scala/com/nvidia/spark/rapids/shims/v2/GpuShuffleExchangeExec.scala:package com.nvidia.spark.rapids.shims.v2
./sql-plugin/src/main/311db/scala/org/apache/spark/rapids/shims/v2/GpuShuffleExchangeExec.scala:package org.apache.spark.sql.rapids.shims.v2
./sql-plugin/src/main/301db/scala/com/nvidia/spark/rapids/shims/v2/GpuShuffleExchangeExec.scala:package com.nvidia.spark.rapids.shims.v2
./sql-plugin/src/main/301until310-nondb/scala/com/nvidia/spark/rapids/shims/v2/GpuShuffleExchangeExec.scala:package com.nvidia.spark.rapids.shims.v2

This is working today because this class isn't referenced only in code within the specific shim which uses the appropriate package path. If we ever attempt to commonize code that references this class it will break the build since common code can't directly reference the shimmed class via a single package path (which is the main advantage of the v2 style shim).

[res-life]

@jlowe

One issue is the inconsistent of package and path of GpuShuffleExchangeExec in 312db directory

./sql-plugin/src/main/312db/scala/org/apache/spark/rapids/shims/v2/GpuShuffleExchangeExec.scala

package: org.apache.spark.sql.rapids.shims.v2
path:     org/apache/spark/rapids/shims/v2

Another issue:
org.apache.spark.MapOutputStatistics can't be accessed from outside of [spark] package

private[spark] class MapOutputStatistics

Is the following solution ok?
create one class named "GpuShuffleExchangeExecBase" in org.apache.spark package and create "GpuShuffleExchangeExec" which extends the base class in the expected package org.apache.spark.sql.rapids.shims.v2

[jlowe]

create one class named "GpuShuffleExchangeExecBase" in org.apache.spark package

There's already an org.apache.spark.sql.rapids.execution.GpuShuffleExchangeExecBase which isn't shimmed, but I don't think that is going to help here. The Databricks version, and only the Databricks version, needs to override a method that has MapOutputStatistics in the method signature. That means there's no choice but to have that class be somewhere in the org.apache.spark.... hierarchy to access that class.

I think the proper solution is to move all the GpuShuffleExchangeExec shim v2 classes to org.apache.spark.rapids.shims.v2 so they're consistent. Yes, most implementations won't need to be in that package, but I think it's more important for the class paths to be consistent across shims in the v2 shim approach.