[feature] support post-processors

[JPXKQX]

This PR allows the user to create their own post-processors for the graph. In particular, we provide an implementation of RemoveUnconnectedNodes which may be useful for LAM, where we want to remove the nodes far from the area of interest.

The post-processors are included in the recipe as follows:

nodes:
   ...
edges:
   ...

post_processors:
  - _target_: anemoi.graphs.processors.RemoveUnconnectedNodes
    nodes_name: data
    ignore: cutout_mask # optional
    save_mask_indices_to_attr: indices_connected_nodes # optional

Here, we have 2 optional arguments:

• ignore: It allows the user to ignore some nodes based on an existing attribute. For example, do not drop nodes from inside the limited area, even if they are not connected.
• save_mask_indices_to_attr: It allows the user to store the indices to mask the nodes from the previous graph to the processed graph as an attribute. This may be needed for training/inference using only the points of interest.

[dietervdb-meteo]

There is a bug in GraphDescriptor if one drops unconnected nodes from the graph. In its function get_node_summary() please change

attributes.remove("x")

to

exclude = ["x","indices_connected_nodes"]
attributes = [a for a in attributes if a not in exclude]

Edit: Actually the above is not a proper fix, since "indices_connected_nodes" is a name specified in the config. So it should be replaced with the corresponding config entry or another [more robust?] solution needs to be found.

[dietervdb-meteo]

Similarly to the bug in GraphDescriptor there is one in plotting.prepare.py. In the function get_node_attribute_dims please change

attr_dims = {}
for nodes in graph.node_stores:
    for attr in nodes.node_attrs():
        if attr == "x" or not isinstance(nodes[attr], torch.Tensor):
            continue

to

attr_dims = {}
exclude = ["x","indices_connected_nodes"]
for nodes in graph.node_stores:
    for attr in nodes.node_attrs():
        if attr in exclude or not isinstance(nodes[attr], torch.Tensor):
            continue

Edit: Actually the above is not a proper fix, since "indices_connected_nodes" is a name specified in the config. So it should be replaced with the corresponding config entry or another [more robust?] solution needs to be found.

[dietervdb-meteo]

Could be nicer if in BaseMaskingProcessor the mask is stored as self.mask rather than passed along as an argument between each of the routines. The compute_mask function of subclasses could then set that mask attribute.

[dietervdb-meteo]

Looks good to me now. Thanks Mario!