This fixes the test case added in #8498.

The patch conditionally adds a ColumnarToRowRemovalGuard node that does nothing on top of a

+- ColumnarToRow
   +- FileScan parquet

which is to be cached to avoid this Spark code from removing the C2R, the plan will become:

ColumnarToRowRemovalGuard
+- ColumnarToRow
   +- FileScan parquet [l_orderkey_read#128L] Batched: true, DataFilters: [], Format: Parquet, Location: InMemoryFileIndex(1 paths)[file:/tmp/spark-e732391d-d3f4-45e7-ae2e-d521d7658b01], PartitionFilters: [], PushedFilters: [], ReadSchema: struct<l_orderkey_read:bigint>

which will be treated as regular row-based plan by ColumnarCachedBatchSerializer then be handled with vanilla Spark batch serializer.

Previously, there was an columnar batch type mismatch error because Spark's cache planner removes the top ColumnarToRow but treats the remaining of the plan as vanilla Spark columnar plan.