Let TF-GNN choose between keras or tf_keras consistently with TF 2.15:
both provide Keras 2.15, but it matters which one is used,
because they have separate class hierarchies and global registries.
Along the way, refactor the nested case distinctions of tf_internal.py
into a clear list of supported older TF/Keras versions.
Let TF-GNN choose between keras or tf_keras consistently with TF 2.15: both provide Keras 2.15, but it matters which one is used, because they have separate class hierarchies and global registries.
Along the way, refactor the nested case distinctions of tf_internal.py into a clear list of supported older TF/Keras versions.