feat(aggregation): Add MoDoWeighting

[KhusPatel4450]

Adds MoDoWeighting from Three-Way Trade-Off in Multi-Objective Learning: Optimization, Generalization and Conflict-Avoidance (JMLR 2024).

It's a stateful Weighting[PSDMatrix] implementing the λ-update from Algorithm 2:

• λ_{t+1} = softmax(λ_t − γ·(G·λ_t + ρ·λ_t))

Per the discussion with @PierreQuinton and @ValerianRey on Discord, this follows the official LibMTL implementation which uses softmax rather than the paper's hard simplex projection.

Designed to be composed with autogram.Engine in a two-batch training loop so that MoDo's double-sampling property is preserved (Gramian comes from batch 1; backward uses batch 2).

Test plan

• Unit tests in tests/unit/aggregation/test_modo.py (12 functions, 72 cases — structural, reset, parameter validation, softmax boundary cases, recurrence verification)
• Full unit suite passes: 3098 passed, 66 skipped, 33 xfailed
• ty check passes on _modo.py
• Sphinx doctest: 97 tests, 0 failures
• HTML build clean with -W --keep-going -n
  EOF
  )"

[PierreQuinton]

Looking good, it still needs few changes but once this is merge, I think this makes #676 easier to merge.

[PierreQuinton]

So in the end, this is a softmax. @rkhosrowshahi I think this means that moco is essentially just a composition with this weighting, where essentially you give yy_t to it, and then multiply yy_t by the obtained weights. Is that correct? If yes, I think we should change #676 accordingly.

[PierreQuinton]

/opencode:Plan is the inheritance order correct here?

[KhusPatel4450]

it matched _gradvac.py, and also the warning in docstring for _NonDifferntiable states "Placing this mixin before the primary base will cause it to shadow the primary class's call signature in generated documentation."

So yes, I believe it is

[PierreQuinton]

Thanks!

[PierreQuinton]

For me this is ready, let's wait for @ValerianRey 's review s still.