[FEA] Expose isnull/isnan

[quasiben]

Data is often missing and messy. It would be nice to expose isna and isnull methods on series objects. These methods are often used for during filter operations like those below:

cdf = cudf.read_csv(...)
cdf[cdf.isna()]

cdf[cdf.isnull()]

[kkraus14]

@dillon-cullinan You should be able to tackle this as part of #1126 since I think you already generate the boolean mask.

[beckernick]

It would also be good (for downstream usage and pandas API compatibility) to expose this functionality as a top level function. isna and isnull can be access via pd.isnull and pd.isna.

[beckernick]

Lack of the isnull method is now partially blocking a dask change to permit cumulative aggregation operations in dask-cudf

[randerzander]

@dillon-cullinan are you able to tackle isnull/isna as part of #1126?

@beckernick can you confirm if we need to add special logic for handling strings too?

[beckernick]

In terms of blocking cumulative aggregations, since the cumulative aggregations are are only operating on numeric columns these methods could be implemented at the series level and raise NotImplementedErrors (for now) if the column is a StringColumn.

In general, I think whether we need different logic than the numba kernel work in #1126 depends on whether the StringColumn null mask behaves the same way as a cudf::column. If it does, I think we're fine.

[beckernick]

As an update, this specific issue is not necessarily blocking the dask functionality. Either isnull or notna methods (and top-level functions) would suffice as long as isnull could be used with the negation operator such as ser[~ser.isnull()].