Move empty directory removal to sparseml instead of sparsify

[dsikka]

• Default behaviour involves the user passing in just the data arg and this always results in imagefolder being populated in for the dataset arg, for both custom datasets and datasets where we download public data (this shouldn't be happening/should be separated further in a follow-up PR)
• However, we only want to check the existence of folders for custom datasets, not public datasets we haven't pulled down.
• Therefore, remove the check and error
• This should also just be done in sparseml, since click.Path() is where the folder is created.

Tested locally using the following command:
sparsify.run sparse-transfer --use-case image-classification --data imagenette --optim-level 0.5