Approximately recreating Jia results

This release is code and results for approximately recreating results from Xiaowei Jia's original model. There are a few notable differences:

1. Although the model was trained on the subset of the river network (42 segments), predictions were made for the entire network (456 segments)
2. The model was trained using sequential data sequences one-year in length. XJ trained the model with a half year overlap in the batches. So he trained on twice as many batches (half of them were the same as the ones here, half of them were offset by half a year).
3. In XJ's model there were different weights for the outputs of the pretraining and the finetuning. In our implementation, there is only one set of weights which is the same between pretraining and finetuning.