Improve interface routines

[nbren12]

I need functions that make it easy to call the neural network with

• [ ] dicts of numpy arrays
• [ ] Stacked numpy arrays
• [ ] xarray objects
• [ ] dicts of torch arrays

To do this, I should follow a hub and spoke model. The hub will be a common data-structure (e.g. dict of torch arrays) and the spokes will convert data-structures to this central data structure. Then, I should only need to write one interface routine for the central data structure, and then provide translation routines from xarrays to this data structure.

• this central data format will be a dict of torch tensors with shapes (time, y, x, z) where c can be either 1 or the number of vertical points. All tensors will need to be broadcast-able to this shape. This won't work because individual points are selected in x and y and the concatenated along one batch dimension.
• What about dict of tensort with shape (time, batch, z)? This is what I currently have.

[nbren12]

done. uwnet.interface offers interfaces for numpy, xarray, and torch.