feat(aggregation): Add Composition class and << operator

[powerofaisinstudy-debug]

Summary

Closes #775.

This PR adds support for composability of aggregators and matrix transformations:

• Implemented Composition in src/torchjd/aggregation/composition.py.
• Overloaded __lshift__ operator in Aggregator (src/torchjd/aggregation/_aggregator.py).
• Added unit tests in tests/aggregation/test_composition.py.

@PierreQuinton Here is an initial draft implementation ready for review whenever you have time!

[powerofaisinstudy-debug]

@PierreQuinton Here is an initial implementation ready for review whenever you have time!

Could you please add the cc: feat label to this PR so the format-title CI check passes? I don't have permission to edit labels. Also, the Check Links workflow is timing out on an external journal link in the existing repository files (comptes-rendus.academie-sciences.fr).

[PierreQuinton]

Thanks for the PR. Could you adapt the code to match what is done in https://github.com/SimplexLab/TorchJD/blob/main/src/torchjd/autojac/_transform/_base.py ?

Mainly:

• Put Composition in the _aggregator_bases.py file.
• Make Composition take exactly one aggregator and one function, and type them correctly (the typing was provided in [Feature Request] Add support for custom conflict-resolution hooks in gradient aggregation #775 ).
• Add a compose function with the appropriate typing, and set __lshift__ = compose.
• Implement forward rather than __call__ in Composition.
• Implement __str__ rather than __repr__, similar to

  TorchJD/src/torchjd/autojac/_transform/_base.py

  Line 75 in 30db018

  return str(self.outer) + " ∘ " + str(self.inner)

• Add tests

[PierreQuinton]

@ValerianRey Can this be implemented using a hook? So basically, can a hook change the value of the input matrix, or is it just an observer?

[ValerianRey]

@ValerianRey Can this be implemented using a hook? So basically, can a hook change the value of the input matrix, or is it just an observer?

Yes. Also, this PR doesn't work because the output of the first aggregator in the composition will be a vector, not a matrix. We can't compose aggregators since their domains don't match.