[FEA] Add large strings support in CSV writer

[GregoryKimball]

Is your feature request related to a problem? Please describe.
The libcudf CSV writer throws a std::overflow_error when trying to file a dataframe or a chunk that exceeds 2.1B characters.

Describe the solution you'd like
There are a few options:

1. fix the overflow error in the CSV writer implementation. This would be a libcudf change. Since the CSV writer is fairly simple and relies on libcudf strings API, there may be a straightforward solution.
2. Change the default chunksize to libcudf or cuDF from std::numeric_limits<size_type>::max(); to something based on the dataframe size, e.g. chunksize = len(df) // (df.memory_usage(deep=True).sum() / 500_000_000.)

Additional context
As far as fixing the root cause, here is a quick repro:

    df = cudf.DataFrame({'text': ['a'] * 1000})
    df['text'] = df['text'].str.repeat(1000)
    data_1gb = cudf.concat([df] * 1000, ignore_index=True)
    df = cudf.concat([data_1gb] * 3, ignore_index=True)
    df.to_csv('/raid/tmp.csv')

which throws:

  File "csv.pyx", line 547, in cudf._lib.csv.write_csv
OverflowError: Writing CSV file with chunksize=3000000 failed. Consider providing a smaller chunksize argument.

but I couldn't immediately find where the std::overflow_error is getting thrown in io/csv/writer_impl.cu.

As far as chunksize, I did some analysis of chunksize, and using smaller chunks of ~100 MB shouldn't have a significant performance impact.

[davidwendt]

I know where this is occurring and will look into this.