Optimise PyTorch MLP

[mastoffel]

Currently, the PyTorch MLP tends to be a bit worse than the scikit-learn MLP on different datasets. Would be good to find out why, e.g. could it be the optimiser (lbfgs works well for small datasets but isn't implemented in PyTorch).

[kallewesterling]

Any updates on this issue @mastoffel and @bryanlimy? It looks like the only one pending from our sprint planning meeting currently (beside release)!

[kallewesterling]

@mastoffel just to clarify, does #205 not resolve #192?

[mastoffel]

@kallewesterling only partially!