[Feature Request]: Remove need for tf.function in TF ModelHandler

[damccorm]

What would you like to happen?

Right now, to use the Tensorflow model handler you need to take in a string tf.example. To support models that don't do this and take in something like a numpy array of floats, we currently recommend adding a tf.function (see https://github.com/apache/beam/blob/c36077a96a4691ca9000c8b243847ac6c7f82a74/examples/notebooks/beam-ml/run_inference_tensorflow.ipynb). We should automatically support this use case, and ideally do so more generally.

Issue Priority

Priority: 2

Issue Component

Component: run-inference

[damccorm]

Completed by #25366