[BUG] legacy struct cast to string crashes on a two field struct

[gerashegalov]

Describe the bug
In Spark 3.0.x or in Spark 3.1.+ with spark.sql.legacy.castComplexTypesToString.enabled=true queries from an RDD source may crash with

java.lang.AssertionError:  value at 15 is null
        at ai.rapids.cudf.HostColumnVectorCore.assertsForGet(HostColumnVectorCore.java:228)
        at ai.rapids.cudf.HostColumnVectorCore.getUTF8(HostColumnVectorCore.java:355)
        at com.nvidia.spark.rapids.RapidsHostColumnVectorCore.getUTF8String(RapidsHostColumnVectorCore.java:177)
        at org.apache.spark.sql.vectorized.ColumnarBatchRow.getUTF8String(ColumnarBatch.java:220)
        at org.apache.spark.sql.catalyst.expressions.GeneratedClass$SpecificUnsafeProjection.apply(Unknown Source)
        at org.apache.spark.sql.catalyst.expressions.GeneratedClass$SpecificUnsafeProjection.apply(Unknown Source)
        at scala.collection.Iterator$$anon$10.next(Iterator.scala:459)
        at org.apache.spark.sql.execution.SparkPlan.$anonfun$getByteArrayRdd$1(SparkPlan.scala:346)
        at org.apache.spark.rdd.RDD.$anonfun$mapPartitionsInternal$2(RDD.scala:898)
        at org.apache.spark.rdd.RDD.$anonfun$mapPartitionsInternal$2$adapted(RDD.scala:898)
        at org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:52)
        at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:373)
        at org.apache.spark.rdd.RDD.iterator(RDD.scala:337)

Steps/Code to reproduce bug
Minimum repro:

import sys
sys.path.append(r'/path/to/localgit/spark-rapids/integration_tests/src/main/python')

import pyspark.sql.functions as F
import pytest
from data_gen import *

key_data_gen = StructGen([
        ('a', IntegerGen(min_val=0, max_val=4)),
        ('b', IntegerGen(min_val=5, max_val=9)),
    ], nullable=False)
val_data_gen = IntegerGen()
df = two_col_df(spark, key_data_gen, val_data_gen)

# For Spark 3.1+
spark.conf.set('spark.sql.legacy.castComplexTypesToString.enabled', True)
df.select(df.a.cast(StringType())).filter(df.b > 1).collect()

GPU plan exhibiting the crash:

21/04/30 17:37:16 WARN GpuOverrides: 
*Exec <ProjectExec> will run on GPU
  *Expression <Alias> cast(a#0 as string) AS a#34 will run on GPU
    *Expression <Cast> cast(a#0 as string) will run on GPU
  *Exec <FilterExec> will run on GPU
    *Expression <And> (isnotnull(b#1) AND (b#1 > 1)) will run on GPU
      *Expression <IsNotNull> isnotnull(b#1) will run on GPU
      *Expression <GreaterThan> (b#1 > 1) will run on GPU
    !NOT_FOUND <RDDScanExec> cannot run on GPU because no GPU enabled version of operator class org.apache.spark.sql.execution.RDDScanExec could be found
      @Expression <AttributeReference> a#0 could run on GPU
      @Expression <AttributeReference> b#1 could run on GPU

Expected behavior
Cast should work the same as on CPU

Environment details (please complete the following information)
local REPL is sufficient to reproduce

Additional context
bug found while working on #2274 .

Interestingly saving the synthetic df to parquet and reading it back yields the correct result without a crash with the FileSourceScanExec plan:

*Exec <ProjectExec> will run on GPU
  *Expression <Alias> cast(a#38 as string) AS a#42 will run on GPU
    *Expression <Cast> cast(a#38 as string) will run on GPU
  *Exec <FilterExec> will run on GPU
    *Expression <And> (isnotnull(b#39) AND (b#39 > 1)) will run on GPU
      *Expression <IsNotNull> isnotnull(b#39) will run on GPU
      *Expression <GreaterThan> (b#39 > 1) will run on GPU
    *Exec <FileSourceScanExec> will run on GPU