FR: Switch to torchvision for MNIST Dataset loading

[stroblme]

Is your feature request related to a problem? Please describe.
Currently there is a dependency on tensorflow only for loading the MNIST dataset.
Furthermore the MNIST dataset has to be loaded manually into the appropriate folder.

Describe the solution you'd like
Use the dataset handler available with torchvision for loading the MNIST dataset.

Describe alternatives you've considered
Alternatively we could use a manual approach where we introduce a custom Dataset for Kedro that inherits from the Abstract Dataset class and handles the loading and providing of MNIST data internally.
Although this approach might be more in the spirit of Kedro, it is also more prone to errors and potentially more time consuming compared to the well established code from torchvision that would allow an almost drop-in replacement with the existing code.

Additional context
N/A

[stroblme]

Postponed as not urgent and switching to torchvision requires quite some restructuring of code as torch returns full dataloader object that would break the current preprocessing pipeline