Add GRU-D model for mortality predition.

[tracy030115]

Contributor: Tracy Zhang (tracyz3@illinois.edu)
Contribution Type: Model (Reproducibility and ablation)
Paper: https://arxiv.org/abs/1908.00690
Description: Implemented GRU-D (Gated Recurrent Unit with Decay), a RNN designed for irregularly sampled clinical time series with missing values. GRU-D extends the standard GRU with two decay mechanisms: input decay and hidden state decay. The model compared Raw, CUI, and Clinical feature representations along with hyperparameter tuning and ablation on the two decay mechanisms.
Files to Review:

• pyhealth/models/grud.py
• pyhealth/models/__init__.py
• tests/core/test_grud.py
• docs/api/models/pyhealth.models.GRUD.rst
• docs/api/models.rst
• examples/mortality_prediction/mimic3_mortality_grud.ipynb