This PR extends some of the current functionality:

• It refactors the current node builders, adding a class attribute hidden_attributes to define the attributes of the class that are stored in the graph during creation, and are deleted after the graph is built. These attributes are stored in their respective nodes with a prefixed "_".
• We remove the CutOutZarrDatasetNodes. Now the cutout is passed directly to ZarrDatasetNodes. We extract the cutout mask definition from the node builder into a special node attribute builder, CutOutMask.
• 2 new node attribute builders, MissingZarrVariable and NotMissingZarrVariable are added. They create a boolean node attribute corresponding to whether or not a particular variable is NaN in the first sample of the dataset. This will allow the user to define graphs only where ocean variables are defined.

Below I show an example of how the LAM config file is changed:

Before

nodes:
  data:
    node_builder:
      _target_: anemoi.graphs.nodes.CutOutZarrDatasetNodes
      lam_dataset: ${hardware.paths.data}/${hardware.files.dataset}
      forcing_dataset: ${hardware.paths.data}/${hardware.files.forcing_dataset}
      thinning: 25

Now

nodes:
  data:
    node_builder:
      _target_: anemoi.graphs.nodes.ZarrDatasetNodes
      dataset: 
        cutout:
        - dataset:${hardware.paths.data}/${hardware.files.dataset}
          thinning: 25
        - dataset: ${hardware.paths.data}/${hardware.files.forcing_dataset}
        adjust: all
    attributes:
      cutout:
        _target_: anemoi.graphs.nodes.attributes.CutOutMask