Specialized epistatic variants

[jgallowa07]

For the gcreplay experiments, we're seeing common mutations which, when paired with other certain mutations, seem to have strong epistatic effects. We would like our models to have a few extra interaction parameters for these combinations before the latent space.

To do this, I imagine the user can specify the mutation combinations of interest using during model initialization. Because models are initialized with during the prep stage where the BinaryMapDataset is created. We can then add an attribute to the dataset that informs a TorchdmsModel to create the extra parameters before the latent space.

@wsdewitt Had an idea about how the model will do this with submodules. Given a description of this, we'll move forward with specifics here.