fix: handle loss config access for class_similarities

[MohamedAliBouhaouala]

Summary

The following PR fixes the logic for accessing class_similarities in the loss config. Previously, the code assumed a fixed structure, which could break when feature_config.loss was a dict or an object. The updated logic handles both cases safely.
While training my model with class_similarities, I encountered this exception:

 TypeError: argument of type 'SoftmaxCrossEntropyLossConfig' is not iterable

Changes

• Added conditional checks to support both dict and object access styles for feature_config.loss.
• Ensured compatibility with different configuration formats (e.g., YAML or Python objects).

Additional Note

I also modified .pre-commit-config.yaml based on a previous PR 4039 to ensure that pre-commit hooks pass correctly. This change was required to avoid commit failures due to missing staged updates in the pre-commit config.