fix and enhance masked output

[dietervdb-meteo]

What happened?

Output has an extract_lam option that currently is broken (it doesn't get the cutout mask correctly out of the checkpoint). This should be fixed.

At the same time during training one has the option to set an output_mask in the model part of the config. This mask is now also stored in the checkpoint. It thus seems natural to define a more general masked_output to replace extract_lam. The idea would be that one can specify in the config which mask to apply to crop the output, but that the default will be output_mask (a feature of the model) , rather than the cutout_mask (which is a feature of the dataset).

What are the steps to reproduce the bug?

Run inference on a lam model checkpoint with a config including

output:
  extract_lam:
    output:
       <my_favorite_output_type>

Version

current develop

Platform (OS and architecture)

irrelevant

Relevant log output

File "XXX/lib/python3.12/site-packages/anemoi/inference/outputs/extract_lam.py", line 28, in __init__
    mask = self.checkpoint.load_supporting_array(points)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "XXX/anemoi-auto-test/lib/python3.12/site-packages/anemoi/inference/checkpoint.py", line 216, in load_supporting_array
    return self._metadata.load_supporting_array(name)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "XXX/lib/python3.12/site-packages/anemoi/inference/metadata.py", line 741, in load_supporting_array
    raise ValueError(f"Supporting array `{name}` not found")
ValueError: Supporting array `cutout_mask` not found
2024-12-18 10:15:12 ERROR 
💣 Supporting array `cutout_mask` not found
2024-12-18 10:15:12 ERROR 💣 Exiting

Accompanying data

No response

Organisation

RMI