Add support for different TFDS BuilderConfigs

[dedeswim]

As of now a user can't use a TFDS BuilderConfig different than the default one: the current parser assumes that the dataset is in the form {torch,tfds,...}/dataset, while in the case of TFDS datasets it can be tfds/dataset/builder-type, e.g. tfds/diabetic_retinopathy_detection/btgraham-300 (as in here).

This PR fixes this by taking whatever is split at line 10 after the first element (which is the source of the dataset) and joining it again with /.

In the case there is just one /, then name[1:] will contain just one element, e.g., if we pass tfds/imagenet2012, then name[1:] == ["imagenet2012"] and "/".join(name[1:]) == "imagenet2012", leaving the current behavior unchanged. If instead there are multiple / (e.g. "tfds/diabetic_retinopathy_detection/btgraham-300"), then name[1:] == ["diabetic_retinopathy_detection", "btgraham-300"], and "/".join(name[1:]) == "diabetic_retinopathy_detection/btgraham-300", which is compatible with the tfds format.

[dedeswim]

I expanded a bit the PR description to clarify how and why this change works