LAM rollout

[gmertes]

When doing inference with LAM, the model will output an interior prediction + the boundary conditions.

During rollout, when we feed the output back to the input, we need a way to replace the boundary conditions with new initial condition at the next timestep (e.g.: ERA5).

We could treat this boundary layer as a forcing: input fields that get updated at every tilmestep from disk. Using the cutout as a mask.