[ENH]: mention chunksize argument when to_csv fails

[miguelusque]

Describe the bug
I am facing the following error when saving a CSV file after reading a DataFrame from parquet.

import cudf

df = cudf.read_parquet("10_000_000a_rows.parquet")
df.to_csv("10_000_000a_rows.csv",index=False)

---

RuntimeError Traceback (most recent call last)
/tmp/ipykernel_1285/3527798299.py in
----> 1 df.to_csv("10_000_000a_rows.csv",index=False)

/opt/conda/envs/rapids/lib/python3.9/site-packages/cudf/core/dataframe.py in to_csv(self, path_or_buf, sep, na_rep, columns, header, index, encoding, compression, line_terminator, chunksize, storage_options)
6332 from cudf.io import csv
6333
-> 6334 return csv.to_csv(
6335 self,
6336 path_or_buf=path_or_buf,

/opt/conda/envs/rapids/lib/python3.9/contextlib.py in inner(*args, **kwds)
77 def inner(*args, **kwds):
78 with self._recreate_cm():
---> 79 return func(*args, **kwds)
80 return inner
81

/opt/conda/envs/rapids/lib/python3.9/site-packages/cudf/io/csv.py in to_csv(df, path_or_buf, sep, na_rep, columns, header, index, encoding, compression, line_terminator, chunksize, storage_options)
239 )
240 else:
--> 241 libcudf.csv.write_csv(
242 df,
243 path_or_buf=path_or_buf,

csv.pyx in cudf._lib.csv.write_csv()

csv.pyx in cudf._lib.csv.write_csv()

RuntimeError: cuDF failure at: /workspace/.conda-bld/work/cpp/src/column/column_factories.cpp:82: Column size cannot be negative.

Steps/Code to reproduce bug
The file is too big to share here. I will share the file with @beckernick internally.

Expected behavior
No error.

Environment overview (please complete the following information)
RAPIDS 22.12 using NGC container.

Environment details
DGX-1 Server

[wence-]

To avoid this error, write using a specified number of rows at a time by providing chunksize=N to to_csv.

This error is occuring because although any single column does not reach the max row count in cudf, the way a CSV file is written is for all columns to be converted to string type in a table, and then all of those columns are concatenated together into a single string column containing the data for the entire dataframe (this is then written to file). By providing a chunksize argument, you're requesting that this concatenation happens N rows at a time, which helps to control the total size of the output string column.

[wence-]

Perhaps we should catch this error in cudf and raise a more informative one (suggesting to specify chunksize).