Allow avoiding the download of extra, when loading the train split of SVHN

Is your feature request related to a problem? Please describe.
I am frustrated when I want to load the train split of SVHN on a new server, because it will always also download the extra split, which is huge.

Describe the solution you'd like
It would be nice to have a flag in the load function, like download_split_only, that allows to only download (and prepare) the necessary splits as specified in the split argument of the function.

Describe alternatives you've considered
I might be able to manually download only the data I need, but I would like to avoid this solution.

Additional context
You can see the problem happening in this colab notebook:

import tensorflow_datasets as tfds

ds = tfds.load(
    'svhn_cropped',
    split='train',
    as_supervised=True,
)

Start by reproducing the tfds.load('svhn_cropped', split='train', as_supervised=True) example and inspect the SVHN dataset loading and preparation path. Determine where the requested split is handled, then verify that loading train downloads and prepares only the requested split while preserving existing behavior for other loads.