Feature/mask NaNs in training loss function

[sahahner]

Variables with missing values that are imputed by the imputer should not be considered in the loss.

The NaN masks are prepared in the imputer. The remapper contains a new function to remap the NaN masks from the imputer.

This goes together with PR #72 from anemoi-training.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.84%. Comparing base (f96bcf9) to head (87647b7).

Additional details and impacted files

@@           Coverage Diff            @@
##           develop      #56   +/-   ##
========================================
  Coverage    99.84%   99.84%           
========================================
  Files           23       23           
  Lines         1301     1301           
========================================
  Hits          1299     1299           
  Misses           2        2           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[floriankrb]

This functionality seems to be related to ecmwf/anemoi-training#79
Perhaps the masks.py created by @JPXKQX should move in anemoi-models and a [refactored version of] OutputMask be used here?

[JPXKQX]

I see some similarities between the output masking and the post-processors, but the part that doesn't fit is that the post-processors are only applied at the end of the rollout. Instead, the masking is called not only at the end, but also in between all the rollout steps (to roll out the boundary forcing). So I don't know if it's better to include it as a special post-processor or leave it in the anemoi-training.

I would say that we can do the loss masking here similar to the imputer, but I think the masking should remain in anemoi-training.