(WIP) Recompute def embs between epochs during training

[jasonrute]

This isn't ready to merge yet. I'm making this PR, so others can comment on this approach.

This PR makes it possible to recompute definitions between epochs.

• Pass the data_server directly to the training class
• Use a callback containing a pointer to the data_server to loop over all definitions.
• Copied code from the tfgnn predict code needed to compute definition embeddings

TODO:

• Do full training run with benchmark to see if added complexity is worth it.
• Decide best way to access definition clusters during training. Currently I'm passing a pointer to the data server to the Keras callback.
• Combine large amounts of duplicated code with the tfgnn predict code into one unified place.