ENH: Allow read_csv(engine=pyarrow) to return ArrowExtensionArray-backed columns

[mroeschke]

• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed.
• Added an entry in the latest doc/source/whatsnew/vX.X.X.rst file if fixing a bug or adding a new feature.

The engine="pyarrow" was added in 1.4 and is considered "experimental", so I would like opinions whether it's okay to

1. Just change the return type to ArrowExtensionArray-backed data in 1.5
2. Raise a warning in 1.5 that the return type will change to ArrowExtensionArray-backed in 2.0. I can't immediately think of a nice way to "opt into" this behavior in 1.5 besides calling astype() after the read_csv call.

[lithomas1]

I would be -1 on anything but opt-in. IMO, the behavior should be consistent across all three engines. I know the read_csv is rather bloated with its many keywords, but I still think it would be worthwhile to add a new one, or co-opt use_nullable_dtypes(the proposed new keyword for the nullable EA types), especially given that Arrow EAs are still kinda new/experimental.

In general, I think there is still some discussion to be resolved around when the CSV parser should return extension array types vs numpy types(see #42973 for example).

[mroeschke]

That's a good point and wasn't aware of expanding use_nullable_dtypes. Yeah agreed I think it would be better to leverage that keyword in a fashion like use_nullable_dtypes="pandas"|"pyarrow"|None. Closing this then.