First draft for KFold

[thibaultdvx]

Here is a first draft for KFold.

At initialization, KFold takes one or several CapsDataset (e.g. KFold(dataset) or KFold(dataset_0, dataset_1)).
The public methods are:

• make_splits: to perform the splitting operation. This method should do the job of clinicadl kfold;
• read: to get the splits from a split directory;
• write: after make_splits has been called, to write the splits in a split directory;
• get_splits: returns an iterator of the wanted splits (e.g. kfold.get_splits(splits=[0, 3])).

The splits are returned as a Split object that contains all the relevant information for training, including the training and validation datasets. If multiple datasets have been passed at the initialization of the KFold object, a tuple of Split objects is returned (splits are arranged according to the order of input datasets).

To build the associated training and validation dataloaders, the user must call the build_train_loader and build_val_loader methods of the Split objects. These methods accept common dataloader parameters (batch_size, shuffle, etc.). I chose to separate these two methods because the user may want different parameters for his/her training and validation loaders.

I also felt the need to refactor the get_dataloader and generate_sampler functions.

Beware that it is a draft and nothing is yet tested.