You can convert it to a list with list(ds) and then recompile it as a normal Dataset with tf.data.Dataset.from_tensor_slices(list(ds)). From there your nightmare begins again but at least it’s a nightmare that other people have had before.
Note that for more complex datasets (e.g. nested dictionaries) you will need more preprocessing after calling list(ds), but this should work for the example you asked about.
This is far from a satisfying answer but unfortunately the class is entirely undocumented and none of the standard Dataset tricks work.