Try streaming dataframes with cudf

Ideally we would be able to reuse the streamz.dataframe module with pandas-like libraries, like cudf.

This would require removing explicit type checks like isinstance(obj, pd.DataFrame) with functional checks like "DataFrame" in type(obj).__name__" (or something better, like what is in dask.dataframe.utils.is_dataframe_like) and probably a bunch of other work.

However, if this works then ideally we would be able to reuse a bunch of code, and make streamz.dataframe a bit more generic.

To start, I would probably make a random dataframe of pandas dataframes, and then convert them into cudf dataframes

from streamz.dataframe import Random
import cudf

sdf = Random()
gdf = sdf.map_partitions(cudf.from_pandas, sdf)
gdf.x.sum()

Probably this will break for many reasons, both in streamz.dataframe and in cudf, but iterating on that set of problems might eventually yield something that is generic enough on both sides to work.

[mrocklin]

Also cc @kkraus14, who I suspect will be interested in this

[martindurant]

Closed by #224