to_dataframe dtypes argument could allow functions that take pandas.Series-like object and return new series

[tswast]

Is your feature request related to a problem? Please describe.

Some timestamp-related methods don't quite work with the current string/dtype object logic. For example, I'm having a hard time getting TIME data types to parse as the timedelta64[ns] dtype.

I also think this feature would be useful for extension dtypes like Fletcher (https://github.com/xhochy/fletcher) and GeoPandas (https://geopandas.org/)

Describe the solution you'd like

Look at the dtypes dictionary. If any value is a function / callable, save them for later. Create the initial dataframe using the other dtypes, then follow-up and call the desired transformation functions.

Describe alternatives you've considered

Keep dtypes a simple pass-through to pandas and either:

• Let the developer call their own transformation functions.
• Have second argument for transformations

Additional context

Possibly useful when setting up default dtypes #786 #793

[tswast]

On second-thought, I think keeping dtypes simple is more important. We can have some custom transformations for default dtypes without this feature.

[tswast]

On third-thought, for flectcher something like this could be useful, but the transformation function should take the arrow buffer(s) and return the pandas series. Definitely shouldn't use dtypes for this, though.