The project operator has inconsistent processing logic for projects without selected columns in vanilla spark and ch backend

[zheniantoushipashi]

Describe the issue

When aggregating by a column that does not exist, spark will convert it to aggregate by Literal value of null, Spark will do column Pruning , and the Project operator in the optimized plan will not select any columns. so In the logic of vanilla spark, spark’s project operator returns N rows with no columns. In ch backend, all rows will be discarded in ExpressionActions and an empty data set will be returned, resulting in inconsistencies in the aggregated results of the agg operator

the Physical Plan :

== Physical Plan ==
AdaptiveSparkPlan isFinalPlan=true
+- == Final Plan ==
HashAggregate(keys=[[null]#18], functions=[count(1)], output=[count#10L])
+- CHNativeColumnarToRow
+- ShuffleQueryStage 0
+- ColumnarExchangeAdaptor hashpartitioning([null]#18, 1), ENSURE_REQUIREMENTS, false, [plan_id=145], [id=#145], [OUTPUT] Vector([null]:StructType(StructField(a,IntegerType,true)), count:LongType), [OUTPUT] Vector([null]:StructType(StructField(a,IntegerType,true)), count:LongType)
+- RowToCHNativeColumnar
+- HashAggregate(keys=[[null] AS [null]#18], functions=[partial_count(1)], output=[[null]#18, count#20L])
+- CHNativeColumnarToRow
+- *(3) ProjectExecTransformer
+- RowToCHNativeColumnar
+- LocalTableScan [_1#1]

A clear and concise description of what works not as it is supposed to.

How to reproduce

the spark unit test of "SPARK-34713: group by CreateStruct with ExtractValue" in DataFrameAggregateSuite can reproduce it

test("SPARK-347133: group by CreateStruct with ExtractValue") {
spark.conf.set("spark.sql.codegen.wholeStage", false)
val nonStringMapDF = Seq(Tuple1(Map(1 -> 1))).toDF("col")
// Spark implicit casts string literal "a" to int to match the key type.
checkAnswer(nonStringMapDF.groupBy(struct($"col.a")).count().select("count"), Row(1))

val arrayDF = Seq(Tuple1(Seq(1))).toDF("col")
val e = interceptAnalysisException
assert(e.message.contains("requires integral type"))
}