dtypes_backend options:

• "default" (current "mixed" behavior, where most extension dtypes are used but not Float64Dtype or StringDtype) -- edit: might want to check dtype_backend in pandas global options (per API: Setting Arrow-backed dtypes by default pandas-dev/pandas#51433),
• None (all legacy types),
• "numpy_nullable" corresponds to existing behavior but also nullable floats (feat: Use nullable Float64Dtype to allow NULL and NaN to be represented in the same Series when dtype_backend="numpy_nullable" #14483) and strings (somewhat related: ENH: option efficient memory use by downcasting python-bigquery-pandas#275 (comment)).
• "pyarrow" (all Arrow dtypes)

Implementation note: we probably need to call to_arrow for all but default, where to_dataframe is still OK.

See also: pandas-dev/pandas#51853