The nilearn datasets should have more information on their processing in their docstrings, though this could certainly be expanded !
The majority have undergone some level of preprocessing, though relatively few have been run through fMRIPrep specifically (the main exception I can think of is the fetch_development_fmri dataset).
For the fetch_miyawaki2008 dataset in particular, it’s had some level of processing (see: it’s already masked), and if you’re planning to run a similar analysis as the example you linked, I would definitely recommend preprocessing first !
This dataset background is specific to this dataset; in general, you’d likely want to put the anatomical template used for normalization as a background. For example, Nilearn ships the MNI152 template, which is commonly used for anatomical normalization. Indeed, many functions such as plot_stat_map automatically put the MNI152 template as a background because of how commonly it’s used within the community !
If you’re preprocessing with fMRIPrep, though, you may notice that there are in fact multiple MNI152 templates that you can normalize to. You can grab the specific one that your data has been aligned to from TemplateFlow.