Remove tests for preprocessing inside a functional model

[mattdangerw]

This removes all testing for using string dtypes inside the call graph of a functional model. This will never work in a multi-backend fashion because torch and jax deliberately lack string dtypes.

There is really only one case where adding tokenization into the call graph of a model makes sense--inference in tensorflow. And for that use case, we should switch to recommending an export flow (like tf.keras.export.ExportArchive()). Ditching these tests will simplify the code base, speed up testing, and allow us to focus on workflows that apply to all backends.

Instead we, can focus on the follow options for applying tokenization:

1. Automatically, via a Task and tf.data.
2. Manually, with tf.data.
3. Manually, in eager mode outside of a model.
4. Inside an exported tf.function.

For the export flow, it would look something like:

export_archive = keras.export.ExportArchive()
export_archive.track(classifier)
export_archive.add_endpoint(
    name="serve",
    fn=lambda x: classifier(classifier.preprocessor(x)),
    input_signature=[tf.TensorSpec(shape=(None), dtype=tf.string)],
)
export_archive.write_out("serving_model")

Plenty of planning will still need to be done on export() flows for keras_core, but that's probably out of scope for this PR.

[mattdangerw]

/gcbrun

[jbischof]

Thanks!