pyarrow offset overflow for datasets with large strings

[rasdani]

the dataset.maps in Environment cause pyarrow offset overflow for datasets with large strings, e.g. lcbv5 subset of agentica's DeepCoder dataset.

you can test by running the main of deepcoder env from this commit.
https://github.com/PrimeIntellect-ai/prime-environments/blob/6e2ca026927f5c3bd48c987d1445b0260a26eea8/environments/deepcoder/deepcoder.py#L136-152

fix that worked for me: mapping "prompt" myself and limiting writer_batch_size=16.