Consider switching to keras as engine

[NickleDave]

Keras 3.0 is out and allows tensorflow/torch/jax as backends.

This would let us

• vendor more models instead of re-implement them (user would need to install appropriate backend as needed): e.g., DAS, DAN, etc
• not have to deal with pain points of lightning that I am now running into
• be able to use jax for ops that are still not painful/not possible in torch/tensorflow

I am sure there are similar pain points to using keras with multiple back-ends but it is a simple API that would make it easier for users to implement their own models without having to grok the abstraction I came up with. I would need to test it in a branch and see what breaks