Multi-task CV and various fixes

[andrrizzi]

Description

1. Implement MultiTaskCV. Because the class dynamically inherits from the main CV, this doesn't work if the main CV uses __slots__ but I thought this a rare case that we could avoid supporting.
2. Fix FastDictionaryLoader.get_stats().
   • By default, when there are multiple datasets, it computes the stats for all of them unless the dataset_idx argument is given.
   • I've also removed the batch iteration to simplify a bit the code for the multidataset case and because DictionaryDatasets already have everything in memory anyway so this should be slightly faster than doing a for loop.
3. Extended AutocorrelationLoss and FisherDiscriminantLoss to multivariate features. I've also added a Lorentzian regularization to FisherDiscriminantLoss to avoid it to explode.
4. I've fixed a weird bug in the VariationalAutoencoderCV that resulted in wrong torchscript models.
   • Torchscript complained that I was splitting the output of the encoder into two (for the mean and variance) so I added two separate linear layers mean_nn and logvar_nn.
   • I also modified FeedForward to support a last_layer_activation argument so that we can configure easily the encoder to have an activation layer also to the output layer (which is then fed to mean/logvar_nn).
5. Various fixes to PEP8 style, documentation, and tests.

Status

• Ready to go

[codecov]

Codecov Report

Merging #46 (54ef4b9) into lightning (6b5ac84) will not change coverage.
The diff coverage is 0.00%.