Allow computer vision task models to run without tf.data

[mattdangerw]

First step (of a few), to slowly relax our reliance on tf.data for preprocessing.

Our text models are more heavily reliant on tf.data because of the tf-text dependency. Our image models do not have this constraint.

We could try to allow running preprocessing without tf.data when running on the torch and jax backends. To do so, we would need to stop always converting to a tf.data.Dataset in our pipeline model helper here and find a way to still apply preprocessing to the iterator efficiently.

[mattdangerw]

Contributions are welcome here, but this is a fairly abstract problem that would need some scouting out first. We could try to leverage Keras' DataAdapter here, I'm not sure how to best iterate over the dataset and apply the Keras layer. This is probably something best prototyped for a number of vision tasks first (classification, detection, segmentation).