[Train,Tuning] Support Keras 3 Parallel Training and HP tuning

[ghsanti]

Description

The current examples for Keras/TF do not work, presumably because of Keras 3 which is integrated by default since TF==2.16.

One can still switch to Legacy Keras, but this wouldn't make sense, Keras 3 is interesting and has many new features.

Keras 3 seems to have its own multibackend (only supports JAX from what I can understand) distributed framework..

I wonder whether Keras 3 will be supported here, now that it's becoming independent of the backend?

Use case

There are multiple use cases; the main one is that Keras has a beautiful API that makes writing Networks less verbose, plus one can change backends now.