Update cell classification network

[Adam Tyson]

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

• #304
• #303
• #301

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

• #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:

• #298
• #352
• #353
• #354
• #355
• #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)