Pyarrow strict checks might lead to errors when calling `df.to_parquet`

[andreatgretel]

Values in LLM columns can be converted to different types after generation, say float/dict. This leads to errors when calling df.to_parquet.
This can be particularly tricky for nested columns - for instance, in structured generation, fields being converted to different types would also lead to errors.

A possible solution is falling back to strings everywhere, in such cases.

[johnnygreco]

I think this issue will come into play specifically with the create method, which writes parquet files. It is directly related to #128 – we solved this issue, but I think that only stopped such errors in preview mode (these were due to the dataset analysis creating schemas rather than writing files)