Update cell classification network

[adamltyson]

For many reasons the existing cell classification network (and associated helper functions) is suboptimal:

• Improve efficiency of cube extraction from dask array #304
• Batch size does not properly affect cube generation generator #303
• Improving classification performance. #301

The use of TensorFlow is also causing headaches:
• #293

• Update to Python 3.11 #266

I think now is the time to fix this. Previously I've raised brainglobe/cellfinder-core#220 and brainglobe/cellfinder-core#221. But essentially my idea is to update the code (refactoring as we go) to Keras 3.0 which is now out.

It would also be cool to address this in a way that would support:

• 2D data support #298
• [Feature] Make background channel optional #352
• [Feature] Need an intelligent way of picking batch size. #353
• [Feature] Reduce overfitting when training #354
• [Feature] Freeze upper layers of the network for retraining. #355
• Classification + Analysis step Inefficiencies #356

[sfmig]

First steps (chat with @IgorTatarnikov)

• migrate to Keras 3.0
  • verify accuracy of results (via tests)
  • verify no performance differences (via time/memory benchmarks)