feat(aggregation): Add CRMOGMWeighting

[KhusPatel4450]

• Adds CRMOGMWeighting, a stateful Weighting modifier from Conflict-Reduction Multi-Objective Gradient Methods (NeurIPS 2022)
• Wraps any existing Weighting and smooths its output with an EMA: λk = α·λ{k-1} + (1−α)·λ̂_k
• Generic over the input type so it composes correctly with both WeightedAggregator and GramianWeightedAggregator
• Stateful via the Stateful mixin; reset() restores uniform initial weights

Tests:

• uv run pytest tests/unit/aggregation/test_cr_mogm.py -v, 92 tests covering EMA recurrence, alpha boundaries, reset, structural checks on both aggregator paths
• uv run pytest tests/unit -q, full regression (2889 passed)
• uv run ty check src/torchjd/aggregation/_cr_mogm.py, passes
• Sphinx doctest, 94 tests, 0 failures

[ValerianRey]

Thanks a lot for the PR! I'm gonna review soon! In the meantime, you can try to get the CI the pass

[KhusPatel4450]

Hello,

Happy to say, all checks have been passed! Glad to have got my first PR as well. Looking forward to feedback

[KhusPatel4450]

All the things addressed:

• Reset propagation: reset() now calls self.weighting.reset() if the wrapped weighting is Stateful.

• device/dtype/m from weighting output: forward() now calls self.weighting(stat) first and reads everything from lambda_hat, not from stat.

• Removed repr and the corresponding test_representations test.

• Removed _state_key: _ensure_state now checks shape/dtype/device directly off _lambda.

• Added test for reset() propagation using GradVacWeighting as the inner stateful weighting.

Still open:

• Initial weight strategy (uniform 1/m vs first weighting output)

• Type checking failure

[KhusPatel4450]

Hello I updated the code with the changes that were requested with these two commits, its just that 2nd commit has the similified version and the raise on shape change in CRMOGMWeighting._ensure_state

[ValerianRey]

/opencode:Plan Please review this with a lot of precision.

[ValerianRey]

Ready to merge IMO @KhusPatel4450 @PierreQuinton

To summarize, I:

• Changed the default alpha to 0.9
• Added a initial_weights optional param for lambda_0, and changed the default lambda_0 to be lambda_hat_1 if this parameter is None
• Improved the docstrings and made them doctested
• Removed a remaining .detach() (that would prevent differentiation if user wants it). Note that if the wrapped weighting is NonDifferentiable, the lambda_hat will already be detached anyway, and if the user aggregates a matrix that is detached (as in 99.9% of the cases) this will have no effect. So really this change will not make any difference except in the niche case where a user really wants to differentiate through CR-MOGM.

[PierreQuinton]

LGTM.