Fix callbacks for pre-trained model, use legacy optimizer

[bfhealy]

This PR fixes an error that was raised when running additional training on pre-trained models. The callbacks metadata for the pre-trained model is not saved and so cannot be loaded in like the rest of the model components. Thus, there is now a separate set_callbacks method in nn.py which is called when loading in a pre-trained model.

Additionally, this PR adds compatibility with the latest tensorflow-macos (2.11), which raises an error when accessing tf.keras.optimizers.Adam. The replacement, tf.keras.optimizers.legacy.Adam, is also included in the previous tensorflow-macos version (2.10), so this change will not break scope on environments set up in September 2022 and onward.

(On the topic of tensorflow versions and M1 Macs, see also the updates in #125 pertaining to GPU support.)

[mcoughlin]

Nice! LGTM!

[bfhealy]

@mcoughlin I forgot that for machines other than M1 Macs, we require tensorflow < 2.6, explaining the test failure. I could either remove this limit on the tf version or set up a version-checking if-else in the nn.py code.

[mcoughlin]

@bfhealy I think removing the limit is fine assuming everything works?

[bfhealy]

@mcoughlin Sounds good, I'll remove the limit and see how the tests run.