[BUG] unix_timestamp on GPU returns invalid data for bad input

[andygrove]

Describe the bug

The unix_timestamp function should either return null values or throw an exception for values that cannot be parsed given the specified format (depending on the value of spark.sql.legacy.timeParserPolicy).

However, the function currently returns incorrect data.

Steps/Code to reproduce bug

testSparkResultsAreEqual("Test unix_timestamp mix of valid and invalid inputs",
  spark => {
    import spark.implicits._
    Seq(
      "31/12/1999",
      "31-12-1999",
      "12/31/1999",
      "12-31-1999",
      "1999-12-31",
      "1999/12/31",
    ).toDF("c0")
  },
  conf()) {
  df => df.withColumn("c1", unix_timestamp(col("c0"), "yyyy-MM-dd"))
}

Running on the GPU and on the CPU did not match 
CPU: ArrayBuffer([31/12/1999,null], [31-12-1999,null], [12/31/1999,null], [12-31-1999,null], [1999-12-31,946598400], [1999/12/31,null])

GPU: ArrayBuffer([31/12/1999,-61206796800], [31-12-1999,-61206796800], [12/31/1999,-61806412800], [12-31-1999,-61806412800], [1999-12-31,946598400], [1999/12/31,946598400])

Expected behavior
Behavior should match Spark on CPU.

Environment details (please complete the following information)
N/A

Additional context
N/A

[andygrove]

cuDF now has an is_timestamp function that we could use to filter out invalid values.