FR: support units other than nanoseconds

[tswast]

Currently, one can construct a DateArray with minimal overhead if the dtype is datetime64[ns], but if it is any other datetime64 dtype, it fails. I suspect TimeArray suffers a similar problem.

    @pytest.fixture
    def data():
>       return DateArray(
            numpy.arange(
                datetime.datetime(1900, 1, 1),
                datetime.datetime(2099, 12, 31),
                datetime.timedelta(days=13),
                # dtype="datetime64[ns]"
            )
        )

tests/unit/test_date_compliance.py:35: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
db_dtypes/core.py:49: in __init__
    values = self.__ndarray(values)
db_dtypes/core.py:57: in __ndarray
    return numpy.array([cls._datetime(scalar) for scalar in scalars], "M8[ns]",)
db_dtypes/core.py:57: in <listcomp>
    return numpy.array([cls._datetime(scalar) for scalar in scalars], "M8[ns]",)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

scalar = numpy.datetime64('1900-01-01T00:00:00.000000'), match_fn = <built-in method match of re.Pattern object at 0x1219b4d40>

    @staticmethod
    def _datetime(
        scalar,
        match_fn=re.compile(r"\s*(?P<year>\d+)-(?P<month>\d+)-(?P<day>\d+)\s*$").match,
    ) -> Optional[numpy.datetime64]:
        # Convert pyarrow values to datetime.date.
        if isinstance(scalar, (pyarrow.Date32Scalar, pyarrow.Date64Scalar)):
            scalar = scalar.as_py()
    
        if pandas.isna(scalar):
            return None
        elif isinstance(scalar, datetime.date):
            return pandas.Timestamp(
                year=scalar.year, month=scalar.month, day=scalar.day
            ).to_datetime64()
        elif isinstance(scalar, str):
            match = match_fn(scalar)
            if not match:
                raise ValueError(f"Bad date string: {repr(scalar)}")
            year = int(match.group("year"))
            month = int(match.group("month"))
            day = int(match.group("day"))
            return pandas.Timestamp(year=year, month=month, day=day).to_datetime64()
        else:
>           raise TypeError("Invalid value type", scalar)
E           TypeError: ('Invalid value type', numpy.datetime64('1900-01-01T00:00:00.000000'))

[tswast]

Looks like we might get some of this for "free" with pandas 2.0. Somehow the actual datetime.date values are equal, even with extreme values post-pandas 2.0. See: googleapis/python-bigquery-pandas#627

Edit: Not quite free. We need to be careful that the numpy.datetime64 objects we create are all of uniform units. Might have to refactor, as _datetime will need to get some units passed to it.