allow missing layers in models when loading checkpoint

[kawaho]

This PR allows users to load checkpoints with missing layers compared to the model defined so one can load pre-trained weight when adding more layers to the model.

However, optimizer checkpoint is not handled yet because I imagine we most likely will freeze the pre-trained weight and just train on whatever new layers we are adding. Of course, this feature can be extended to the optimizer checkpoint as well (with more work).

[jpata]

Can you also check pre-commit? I think it might have detected some issues:

https://github.com/jpata/particleflow/actions/runs/12338808099/job/34440017254?pr=380

mlpf/model/training.py:920:13: F841 local variable 'e' is assigned to but never used
mlpf/model/training.py:929:33: F541 f-string is missing placeholders
mlpf/model/training.py:932:33: F541 f-string is missing placeholders

You can also run it locally: https://pre-commit.com/#install

[kawaho]

sorry about that. Should be ok now.

[jpata]

Nice, thank you! Merging.