[BUG] dask_cudf constructors turn None into empty string

[lmeyerov]

Describe the bug

Constructing a dask_cudf Series with some None vals then doing dgs.dropna() on a dask col should drop the None values, but instead they appear to turn into empty strings and thus stick around

Steps/Code to reproduce bug

gs = cudf.Series(['a', 'b', None])

# ### gdf drops None as expected ### #

print(gs.dropna().to_pandas().to_list())  # => ['a', 'b']

# ### dask does not... even with map partitions ### #

with Client('dask-scheduler:8786'):
    print(dask_cudf.from_cudf(gs, npartitions=2).dropna().compute().to_pandas().to_list())
    # => ['a', 'b', '']
    
with Client('dask-scheduler:8786'):
    print(dask_cudf.from_cudf(gs, npartitions=2).fillna(value=np.nan).dropna().compute().to_pandas().to_list())
    # => ['a', 'b', '']

with Client('dask-scheduler:8786'):

    def gdf_dropper(gs):
        return gs.dropna()

    print(dask_cudf.from_cudf(gs, npartitions=2).map_partitions(gdf_dropper).compute().to_pandas().to_list())
    # => ['a', 'b', '']

Maybe dask_cudf is turning None into '' , so a constructor bug?

with Client('dask-scheduler:8786'):
    print(dask_cudf.from_cudf(gs, npartitions=2).compute().to_pandas().to_list())
    # => ['a', 'b', '']

Expected behavior

Dask seems fine here:

with Client('dask-scheduler:8786'):
    print(dask.dataframe.from_pandas(gs.to_pandas(), npartitions=2).dropna().compute().to_list())
    # => ['a', 'b']

Environment overview (please complete the following information)

docker -> mamba -> rapids 0.18

Additional context

Slack: https://rapids-goai.slack.com/archives/C5E06F4DC/p1616723705030300

Blocking: graphistry/pygraphistry#225