[BUG] GpuColumnarToRowExec will always be tagged False for exportColumnarRdd after Spark311

[wjxiz1992]

Describe the bug
when using ColumnarRdd in Spark 3.1.2 (all after), the Shim layer will always use the default value(false) for exportColumnRdd , e.g. https://github.com/NVIDIA/spark-rapids/blob/branch-22.02/sql-plugin/src/main/311until320-nondb/scala/com/nvidia/spark/rapids/shims/v2/Spark31XShims.scala#L457-L459

Steps/Code to reproduce bug
Calling ColumnarRdd in any version after 311, will cause this problem (XGBoost, PCA train):

Expected behavior
There should not be "columnar to row" conversion.

Environment details (please complete the following information)
Spark 3.1.2 Standalone

Additional context
Version before 311 is fine.

[tgravescs]

this looks like its been broken since release 0.2, do we not have a test for this? Do we have an idea of what exactly it causes? If perf issue how much does it affect it.

[wjxiz1992]

Not clear about the XGBoost case. but for PCA case. with 4.6G parquet data , the column is of ArrayType(DoubleType) and the array size if 2048. the perf will drop from 6 seconds to 8 minutes in this case.

without the fix in 3.1.2, training is 8 minutes

but in 3.0.1: 6 seconds