roc_curve: remove nullable type handling when possible

[tamargrey]

There are two ways we could avoid having to explicitly set the numpy dtypes of nullable data after a Series.to_numpy call.

• Pandas stops converting nullable dtypes to object in Series.to_numpy call: API: should to_numpy() by default return the corresponding type, and raise otherwise? pandas-dev/pandas#48891 - then we could just call to_numpy and not have to worry about setting the correct dtype
• Sklearn supports the nullable dtypes in the LabelBinarizer Support nullable pandas dtypes in LabelBinarizer scikit-learn/scikit-learn#25637

If either of the above are completed and released, we can upgrade to that version and make any needed changes to remove nullable type handling from #4019.

[tamargrey]

The related sklearn issues has been fixed, so whenever that work is released and we upgrade to that sklearn version, we should be able to complete this issue.