Description
From the above issues and read_parquet, it appears that generally a use_nullable_dtypes: bool option will be added to read_* functions to allow users to opt into pandas nullable types.
Additionally in 1.5 with ArrowDtype, the nullable type returned could be backed by pyarrow instead of pandas' implementation. This could be advantageous for read_*(engine="pyarrow") capable readers where an option like io.nullable_type="pyarrow" would just preserve the pyarrow object from the pyarrow parsing function instead of converting it to a numpy object.
The proposal would be to add a new global option io.nullable_type="pandas"|"pyarrow" (default="pandas") such that when pd.read_*(..., use_nullable_dtype=True), the nullable backing type would be dictated by the global io.nullable_type setting.