Add filter to config for when trained on overcomplete batches

[dfulu]

Since the _adapt_batch() method we can train a model with batches which are larger than the model requires. i.e. with larger spatial area, longer time horizon, or more inputs (i.e. more NWPs or with/without sat)

However, allowing this means that the config we copy from the batches used to train is overcomplete for the model. Previously we used the workflow of saving the model during train and pushing it to hugging face using the checkpoint_to_huggingface.py` script. But now this means we push a config to huggingface which is overcomplete for the model ans wastes time at inference making bigger batches than required. So we need to edit the config manually somewhere along the way.

This pull request addresses this problem by stripping out parts of the config not required for the model. This is used at the stage when we are pushing the model and its data config to huggingface

[codecov]

Codecov Report

Attention: Patch coverage is 5.88235% with 32 lines in your changes missing coverage. Please review.

Project coverage is 57.77%. Comparing base (5f5d154) to head (ee352d7).
Report is 88 commits behind head on main.

Files	Patch %	Lines
pvnet/models/base_model.py	3.03%	32 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #219      +/-   ##
==========================================
- Coverage   61.39%   57.77%   -3.63%     
==========================================
  Files          28       29       +1     
  Lines        1790     1878      +88     
==========================================
- Hits         1099     1085      -14     
- Misses        691      793     +102     

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

[peterdudfield]

does this variable still et used?

[dfulu]

Yeh it does