Add a dataloader that treats all samples as training

[bentocg]

This would allow us to use the dataset to the fullest extent. We take the best model from hyperparameter search and train it in a combined dataset with all training / validation and test samples.