Split ML functionality dependencies into optional-dependencies

[ns-rse]

TensorFlow is a 600Mb dependency that is now pulled in when installing TopoStats. This can be slow and cumbersome if not on a fast connection (~4 minutes to download on my home connection a supposed 100Mbps connection).

How core is the Tensorflow functionality in the wider community of TopoStats users?

Would it be useful to provide a convenience option of installing without this dependency?

It could be achieved by splitting the dependencies off into their own optional set...

[project.optional-dependencies]
ml = [
  keras,
  tensorflow,
]

And update the documentation to indicate that to get all dependencies for that functionality you should install with pip install topostats[ml].

[ns-rse]

This isn't as simple as suggested as it would require moving the point at which the Unet masking models are loaded in the processing.

[SylviaWhittle]

This would be good - the AFM biomed conference had issues with installing tensor flow dependencies in intel x86 architecture macOS systems