clean up, complete and comment logistic regression example

[marcelluethi]

I restructured and cleaned up the logistic regression example, with the intention of showcasing how simple applications are structured and the core components (like permutations, iterators, jit, ...) are used together.

Instead of relying on LinearMap, which is already a higher level abstraction.

[benikm91]

I overall like the changes.

Why do you put the loss function into the BinaryLR object? For me the model and loss function should be separate concepts. As we can train a model on various loss functions and use a loss function on various models. For me, having a loss function on the model (or score like SKLearn) is a code smell. But for simple example does not matter.

[marcelluethi]

There are two reasons, why combining the loss in the Logistic regression object might be justified:

1. Logistic regression is not only a model but the definition of the loss function is baked into it (if you loose a different loss it's not logistic regression anymore (just another linear model)
2. The loss takes parameters as a parameter. Hence it is tied to the model. If we want to separate the two, we would have to define the loss in terms of predictions and labels, and not parameters and labels.