Adjust InjectorSurrogate module based on LUMETorchModel updates

[pluflou]

This pull request refactors the InjectorSurrogate class to align with changes in LUMETorchModel.

• Refactored the internal state management by introducing the _cache dictionary (as in LUMETorchModel), removing the old _state and update_state methods. The model is initialized with default values (empty set call).
• The model now avoids mutating shared internal state unsafely (self._cache = self.model._cache was aliasing a private attribute).
• Updated the supported_variables property to avoid mutating the wrapped model's metadata and to use a more explicit interface.
• Added helper functions _tensor_to_numpy and _to_python_scalar to safely convert tensors to NumPy arrays and Python scalars.
• Updated examples/injector_surrogate.ipynb to use the new output_beam key instead of beam_output.
• Linting changes to many files.
• Fixed import issue for the YAML model config that's stored in the subtrees. It now temporarily adjusts the paths of the resources at instantiation then deletes the modified YAML. This allows us to keep the surrogate model configs up to date with the upstream repos without having to rewrite the paths.
  • Deleted the old .submodules dir because the naming is confusing (they are not submodules)
  • Added a subtrees/repo_name directory that clarifies what is in there
  • Adjusted model to find the files based on where it's running (CI vs local)