predictions should not all be loaded into memory

[andregraubner]

When predicting, we are just loading all the predictions into memory. This will run OOM fast on smaller machines.

[andregraubner]

When at this, it would be great to give more structure to the predictions (at least when using a high-level abstraction for predicting): E.g. use the file structure convention of the input and output an xarray (holds information about lat/lon etc).
Since the file standard will most likely be a two-group nc file (see #4) where each group can be loaded into an xarray, making the output match the xarray you would get from reading the label group of an input seems reasonable.