Update plotting functions

[niekdejonge]

I started with adding functions for creating the heatmap plots and average prediction per bin plot, but resulted in a large refactoring of the code.

Before we had multiple places in the code where predictions were made, tanimoto scores calculated and losses calculated. These steps are often complex/confusing, since they both have to be the average per inchikey pair (instead of spectrum pair) and the average per bin. Having many different places were these calculations were made and different formats required for plotting, made the code hard to interpret.

This PR is an attempt to reformat the code to make it more harmonized.

@florian-huber
I still had the following questions/choices made that I would like to have your input on:

• I did not implement the weighting function for the validation loss calculation. This can easily be done, but it wasn't really clear to me what the purpose of this was. If it is needed, we can add it. If we don't intend to use it anymore, we could also remove it for the training loss calculations.
• I did remove all code not used in the main training pipeline. Mostly plotting functions that we likely won't use for this paper and will remove from the default plots created. I removed them, since the input required is a bit different than we use for other plots, which could result in using these plotting functions in the wrong way. If we later want to reuse these plotting functions they are still available on older versions on git.
• I did not include the both against both comparison plot. It is also easy to still add, but if we want to add this I think we should think about how we want to implement this. Since now there was no correction done for the number of pos or neg mode spectra available. And spectra from pos and neg with the same inchikeys would still be merged.
• The loss functions during running and loss functions during training are still implemented twice. During training torch is used and during validation pandas is used. Beside this difference, during validation we need to do the step of averaging over the multiple spectrum pairs per inchikey pair. Which we don't want to do during training, which means the mean has to be calculated outside of these methods.