Penalty for straying from process-based equations DO estimates

[jsadler2]

I've more or less thrown in the towel as far as our first implementation of process-guidance goes.

But I have another idea of how we might use the process equations (#47 (comment)) to our advantage:

We start with the metab multi-task. But in the loss function, we add a penalty if the predictions stray from DO estimates made from the process-based equations.