[VL] table scan fallback gets wrong query plan

[FelixYBW]

Backend

VL (Velox)

Bug description

A query has parquet scan with complex data type fallbacked.
If I set spark.sql.parquet.enableVectorizedReader = False and run the query twice, the first one is correct, but the second one wrongly add ColumnartoRow, caused error UnsafeRow cannot be cast to org.apache.spark.sql.vectorized.ColumnarBatch

First query plan:

+- ^ FilterExecTransformer (5)
   +- ^ InputIteratorTransformer (4)
      +- RowToVeloxColumnar (2)
         +- Scan parquet  (1)

The second query plan:

+- ^ InputIteratorTransformer (5)
   +- RowToVeloxColumnar (3)
      +- * ColumnarToRow (2)
          +- Scan parquet  (1)

Similarly if I set spark.sql.parquet.enableVectorizedReader = True, the first run reports error and the second run successed.

Spark version

Spark-3.2.x

Spark configurations

No response

System information

No response

Relevant logs

No response

[FelixYBW]

So there are two issues:

1. wrongly recognized output of vectorized parquet scan as row format
2. some variables isn't reset after a query

[FelixYBW]

1. because customer backport the vectorized parquet scan from spark3.3. ParquetFileFormat.scala in shimlayer needs to be ported.
2. The same issue as [GLUTEN-6151] Reset local property after finishing write operator #6163. isNativeAppliable is set during first query. the local property is used by parquet scan as well. 6163 only reset parquet write.