move dataset objects outside of models

[borauyar]

Fixes #69 except for the GCNNs.

Datasets are not stored within the model (although this doesn't make an impact on either the speed/memory consumption).
The major memory consuming task was Captum Integrated Gradients, which explodes memory consumption, because many models are generated in memory without batching. Now, we use data loaders and compute feature importance in batches, which caps maximum memory usage to close to the amount used during training.

GNNs' feature importance function still needs to be adapted to this.