Test failure on Python 3.8 -- Integer NULL represented as NaN instead of None

[tswast]

========================================================= FAILURES ==========================================================
___________________________ TestReadGBQIntegration.test_should_properly_handle_null_integers[env] ___________________________

self = <tests.system.test_gbq.TestReadGBQIntegration object at 0x7fc296a10a90>, project_id = 'swast-scratch'

    def test_should_properly_handle_null_integers(self, project_id):
        query = "SELECT INTEGER(NULL) AS null_integer"
        df = gbq.read_gbq(
            query,
            project_id=project_id,
            credentials=self.credentials,
            dialect="legacy",
        )
>       tm.assert_frame_equal(
            df,
            DataFrame({"null_integer": pandas.Series([None], dtype="object")}),
        )
E       AssertionError: Attributes of DataFrame.iloc[:, 0] (column name="null_integer") are different
E       
E       Attribute "dtype" are different
E       [left]:  float64
E       [right]: object

tests/system/test_gbq.py:143: AssertionError

[tswast]

I suspect the root cause of this change in behavior is the fact that data in google-cloud-bigquery is now serialized to Arrow before final conversion to DataFrame.

[tswast]

We might want to consider bumping the minimum pandas version up to 0.24.0 and using the "new" nullable integer dtype.

[tswast]

Per the discussion in #242, I think the way forward is to add the dtypes argument and update this particular test to populate it. If dtypes are left unspecified, then it's expected to get different behavior depending on the package versions.