Refactor mappers to allow dynamic behaviour

[JPXKQX]

Is your feature request related to a problem? Please describe.

Currently, mappers in anemoi-models are static: they receive the subgraph (edge indices & edge attributes) at initialisation, storing it as a mapper attribute. This design limits flexibility when applying the model to novel use cases (observations, multi-domain training, ...). Redesigning them to be dynamic, so that the subgraph is passed during forward(), would make mappers more extensible.

Should all mappers be strictly dynamic? Not necessarily. If benchmarks show a significant performance degradation when always passing the subgraph at forward(), we plan to introduce a caching mechanism. This would allow subgraphs to be stored and reused internally through specialised classes.

Current Status

self.encoder = instantiate(  
  model_config.model.encoder, 
  sub_graph=graph[(self._graph_name_data, "to", self._graph_name_hidden)], 
  **encoder_kwargs 
  ...
) 
x_data_latent, x_latent = self.encoder( 
        (x_data_latent, x_hidden_latent), 
        **kwargs 
        ...
) 

Proposed Solution

self.encoder = instantiate(  
     model_config.model.encoder, 
     **encoder_kwargs,
     ...
) 

x_data_latent, x_latent = self.encoder( 
        (x_data_latent, x_hidden_latent), 
        edge_index=graph[(self._graph_name_data, "to", self._graph_name_hidden)].edge_index, 
        **kwargs,
       ...
) 

[ssmmnn11]

the original mappers always received nodes, edges and edge attributes. There will not be a performance impact. I don' think we need caching in any form. This will just depend on how we set up the graph data.