[BUG] Timestamp from GPU ORC reading is different from CPU ORC reading

[ttnghia]

Related: test_read_with_more_columns failed in #6286. As such, when this issue is fixed then that test should also be updated.

See this example to reproduce:

scala> spark.conf.set("spark.rapids.sql.enabled", "false")

scala> import java.sql.Timestamp
import java.sql.Timestamp

scala> val df = Seq(Timestamp.valueOf("1839-12-24 03:58:55.000826")).toDF("v")
df: org.apache.spark.sql.DataFrame = [v: timestamp]

scala> df.coalesce(1).write.mode("overwrite").orc("/tmp/ts.orc")

scala> val df2 = spark.read.orc("/tmp/ts.orc")
df2: org.apache.spark.sql.DataFrame = [v: timestamp]

scala> df2.show(false)
+--------------------------+
|v                         |
+--------------------------+
|1839-12-24 03:58:55.000826|
+--------------------------+


scala> spark.conf.set("spark.rapids.sql.enabled", "true")

scala> val df2 = spark.read.orc("/tmp/ts.orc")
df2: org.apache.spark.sql.DataFrame = [v: timestamp]

scala> df2.show(false)
22/08/12 21:34:09 WARN GpuOverrides: 
*Exec <CollectLimitExec> will run on GPU
  *Partitioning <SinglePartition$> will run on GPU
  *Exec <ProjectExec> will run on GPU
    *Expression <Alias> cast(v#37 as string) AS v#40 will run on GPU
      *Expression <Cast> cast(v#37 as string) will run on GPU
    *Exec <FileSourceScanExec> will run on GPU

+--------------------------+
|v                         |
+--------------------------+
|1839-12-24 03:58:54.000826|
+--------------------------+

[ttnghia]

In particular, the second number is wrongly read. The input second for writing is 55 but read is 54.

[ttnghia]

Turned out this is a bug in libcudf. I've filed the related issue: rapidsai/cudf#11525

[ttnghia]

Other data that can reproduce:

1930-12-24 03:58:55.000826
1950-12-24 03:58:55.000826