Improve demographic transform interface in MedicalFolderDataset
Consider the situation where a researcher specified a demographics file, and they only want to use it for filtering the data, not for using the demographics data inside their ML model. Currently, they are still forced to define a dummy demographics_transform, something like
demographics_transform=lambda x: {}
But this is counterintuitive to them, since they didn't want to use the demographics data for anything!
Internally, we have to do this because we try to convert the demographics to a torch tensor, but a better solution for the researcher is needed