Add temporal architectures

[egillax]

Add temporal architectures to the package. They should include at least:

• A transformer - both including working on visit and feature embeddings
  -- If working on visit embeddings it sums together the feature embeddings for each patient's visit. This will greatly speed up training since the sequnce length will be the number of visits instead of the number of features.

• RNNs like LSTM and GRUs.

• RETAIN - which uses RNNs to compute attentions