add method comparison in training script

[SCiarella]

This PR address #63 by testing which loss function approach we should use between Lux and Optim.jl.

A visual comparison:

Observations:

• Lux is ~25% faster
• Lux is slightly more accurate, however notice that both are bad so further testing of our overall a-posteriori approach is required (maybe this cnn is too weak?)
• Optim can use fancy optimization algorithms like CMA or LBFGS

Conclusions:

• Training with Lux should be our main route
• We can test fancy optimization algorithms with Optim if we put this loss in the package
• Overall, we have no example where the a-posteriori training shows any sensible improvement over random model weights

[luisaforozco]

I am wondering if this validation an plotting should be in this file or in evaluate_posteriori.jl, have you thought about that?

[SCiarella]

Would be nice, however for the comparison we need to access the full training routine, while the idea of evaluate_posteriori is to just load a trained model.

I think that this comparison should go to the explorations repo, but we include both loss functions to src and we show only the Lux one in train_posteriori. What do you think?

[luisaforozco]

Yes, I like the idea:

1. copy the script as it is to explorations (still with the loss defined in it for clarity).
2. adding the code to src/loss/loss_posteriori.jl.

[luisaforozco]

I'm ok with this! I'm merging it.