Introduce common ModynModel Format

[MaxiBoether]

PR #45 initializes a model directory with a dummy ResNet18 class. We need to think about how exactly our models will be implemented, i.e., is it just the class itself, or some meta-wrapper around the concrete implementation (for example, to be able to abstract PyTorch/Tensorflow/...). Then of course, we need to implement some models there.

Maybe it also makes sense to move the models directory into the frontend directory. I will think about that.

[MaxiBoether]

PR #56 introduces a ResNet18 example for PyTorch. This solves previous parts of this issue (implementing a simple model). I think the scope of this issue should be a little different: We should define a general ModynModel class and let other models inherit from this, and build it in a way that (at least for future PRs) support both PyTorch and Tensorflow implementations. PR #56 assumes PyTorch for now, but we will want to generalize a bit here

[vGsteiger]

Is this still an open issue or has it been rediscussed in other issues such as #127 and #111

[MaxiBoether]

It's still somewhat part of a larger issue. Right now we just assume torch. However, in a perfect world, Modyn would support at least Torch/Tensorflow/would be independent as good as possible of torch. For this we probably still need some form of structure in the future that allows us to obtain the correct implementation for each framework. That is what this issue is about

[MaxiBoether]

Pr #294 introduces some functionality models should implement to support Coreset functionality. It was the goal to introduce a ´ModynModel` class in that PR, but did not work out as outlined in the comment here by @francescodeaglio. We should find a way around this and refactor the models into a common parent class.