neuralhydrology/neuralhydrology ? reverse-engineered prompt

Build me a Python library for training neural networks for hydrology research.

I want it to let me train models on different water related datasets without changing the core code every time. The main idea should be that I can define almost everything in a config file, like the dataset, model, loss, optimizer, and training settings, then run training from there. It should be based on PyTorch and feel flexible enough for research, so I can swap in new model ideas or training tricks later.

Please include a clear project structure, documentation, examples I can run, and tests for the main pieces. It should be easy to use for things like streamflow prediction and other hydrology tasks, and it should support modular model components so new research ideas can fit in cleanly. If you need to check current best practices or docs, look them up online.