Modify RowConstructor to work with WithColumn

[suhsteve]

This allows nested Rows with nested StructTypes to be properly unpickled as well as allows Rows to be passed to Udfs by using the WithColumn method.

            _df = _spark
                .Read()
                .Schema("age INT, name STRING")
                .Json($"{TestEnvironment.ResourceDirectory}people.json");

            Func<Column, Column> AgeIsNull = Udf<Row, bool>(
                r =>
                {
                    var age = r.GetAs<int?>("age");
                    return !age.HasValue;
                });

            string[] allCols = _df.Columns().ToArray();
            DataFrame dummyColDF =
                _df.WithColumn("DummyCol", Struct(allCols[0], allCols.Skip(1).ToArray()));
            DataFrame ageIsNullDF =
                dummyColDF.WithColumn("AgeIsNull", AgeIsNull(dummyColDF["DummyCol"]));

Produces the following Schema for the ageIsNullDF Dataframe

{
    "type": "struct",
    "fields": [
        {
            "name": "age",
            "type": "integer",
            "nullable": true,
            "metadata": {}
        },
        {
            "name": "name",
            "type": "string",
            "nullable": true,
            "metadata": {}
        },
        {
            "name": "DummyCol",
            "type": {
                "type": "struct",
                "fields": [
                    {
                        "name": "age",
                        "type": "integer",
                        "nullable": true,
                        "metadata": {}
                    },
                    {
                        "name": "name",
                        "type": "string",
                        "nullable": true,
                        "metadata": {}
                    }
                ]
            },
            "nullable": false,
            "metadata": {}
        },
        {
            "name": "AgeIsNull",
            "type": "boolean",
            "nullable": true,
            "metadata": {}
        }
    ]
}

Benchmark:
HDInsight
D14v2 16 CPU, 112GB Ram
Spark 2.3.2
tpch query 1
dataset size 2.8GB Parquet (generated with db-gen with a 9 scale factor)
master -> 72fdf4834d5d9e7df06efec29233734aaec9445b
PR -> 882cdbe9ac8f657163a6777a72577db45f1b0154

• local[14]

  Iteration	Time(ms) master	Time(ms) PR
  1	90607	83569
  2	93520	88445
  3	95719	91437
  4	89949	78545
  5	94954	85341
  6	96525	77718
  7	98416	87040
  8	72576	69994
  9	78562	77574
  10	84906	89991
  11	77176	79207
  12	97384	72655
  13	99845	81865
  14	103916	87758
  15	96931	92020
  16	98231	102719
  17	101917	100198
  18	103467	105485
  19	83429	93286
  Avg	92527.89	86570.89

• local[7]

  Iteration	Time(ms) master	Time(ms) PR
  1	43714	41642
  2	44725	40108
  3	47259	41202
  4	36858	37197
  5	37770	34779
  6	36003	37000
  7	35127	37738
  8	38190	41163
  9	43216	45113
  10	36987	39170
  11	41157	38587
  12	41458	39658
  13	41671	40793
  14	39415	43389
  15	41639	42874
  16	42248	42231
  17	48151	48480
  18	48973	39301
  19	39647	44853
  Avg	41274.10	40804.10

[imback82]

Can you check the test failures?

[imback82]

LGTM. Great work!