Added architectures (Conv and VAE) and modified loss

Hey @hanikevi ,

I have some changes that should be ready for merge. Would be really nice if you take a look at them. Quick changelog:

• Added convolution architecture (similiar to Cheng et al.),
• added the sanity check we discussed to plot the encoder/decoder weights (fn call currently commented out),
• added modified VAE architecture (with kl_loss computation),
• added modified loss_calculation for VAE training, aligned with your recent changes:
  • if we train VAE there is no pred_offset -> Purpose.RECONSTRUCTION,
  • within the ReconLoss I check the type of the model used,
  • if isinstance(self.model, VAE): I add the kl_loss,
  • then the total_loss is returned a sum of l2_loss and kl_loss.
• added some comments to better initialize the models.

Exceptionally I want to keep the og branch after the merge in terms Lisa has any objections. I'll take care about that in January.

Wish you a nice holiday! 🎄

Closes #64