Fix/SK-1060 | Move mnist data to scaleout public bucket

[Wrede]

This pull request includes significant changes to the data handling and dependencies in the examples/mnist-pytorch/client project. The most important changes involve replacing the use of torchvision with direct data downloads, updating the data loading logic, and modifying the environment dependencies.

Data Handling Improvements:

• examples/mnist-pytorch/client/data.py: Replaced torchvision dataset download with direct data download using requests. This includes setting a random split ID and downloading the corresponding data file from Scaleout's own public bucket.
• examples/mnist-pytorch/client/data.py: Updated the load_data function to dynamically find the correct data split ID and verify the existence of the data file.
• examples/mnist-pytorch/client/data.py: Removed the splitset and split functions, as the data splitting is now handled by the pre-downloaded data files.

Dependency Updates:

• examples/mnist-pytorch/client/python_env.yaml: Removed torchvision from the dependencies and updated numpy versions to ensure compatibility with different platforms and Python versions.

[ahellander]

I think this is a good general improvement, although it reduces the flexibility a bit (max 10 splits possible).

But you need to update the getting started guide as well, e.g. remove the FEDN_NUM_DATA_SPLITS env variable.

Not needed now ,but I think for pedagogical reasons it would also be good to add the option to first download the data split, then start the client. But this could be done later.