Improve performance by constraining data to be read in

If only a subset of model data is to be used then we could improve speed and memory performance by only reading the relavant data in the first place. I think IRIS allows you to this but our Data API doesn't at the moment.