feat(scalarization): Add scalarization package

[ppraneth]

Closes #666.

Summary

Adds a new torchjd.scalarization package providing simple baselines against which aggregators can be compared. Includes:

• Scalarizer: abstract base class, inherits from nn.Module.
• Mean: returns losses.mean().
• Sum: returns losses.sum().
• Constant(weights): returns (weights * losses).sum(). Validates weights.shape == losses.shape at call time.
• Random: combines losses with positive random weights summing to 1 (RLW, Algorithm 2 of arXiv 2111.10603).

All scalarizers accept loss tensors of any shape (including 0-dim) and return a 0-dim scalar.

Design decisions (confirmed with maintainers )

1. Short names mirroring the aggregation package, no Scalarizer suffix.
2. Scalarizer inherits from nn.Module for consistency with Aggregator and Weighting, and to leave room for trainable scalarizers later.
3. No Combiner generalization for now.
4. Arbitrary-shape input, 0-dim output.
5. Constant uses option (a): weights.shape must equal losses.shape.

The Stateful mixin stays in torchjd.aggregation._mixins for now. When the first stateful scalarizer lands, it can be moved to torchjd._mixins so both packages share it. A comment in scalarization/__init__.py records this.

Test plan

• uv run pytest tests/unit/scalarization -W error -v passes.
• uv run pytest tests/unit/scalarization --cov=src/torchjd/scalarization shows full coverage.
• uv run pytest tests/unit -W error passes (3019 passed, 66 skipped, 33 xfailed).
• PYTEST_TORCH_DTYPE=float64 uv run pytest tests/unit/scalarization -W error passes.
• uv run ruff format --check and uv run ruff check pass.
• uv run ty check passes.
• uv run pre-commit run --all-files passes.

[ValerianRey]

Tyvm for this very good PR.
Got a few nitpicks to fix, and then we can merge!

[ValerianRey]

Also @ValerianRey should we define a code owner for the package or leave it as the default (maintainers) and then after maybe a few PR from @ppraneth we can make him code owner?

Yes, I think we can leave it as default and let @ppraneth become code owner if they are interested and add some non-trivial scalarizers in future PRs.

[ppraneth]

Yes, I think we can leave it as default and let @ppraneth become code owner if they are interested and add some non-trivial scalarizers in future PRs.

Thank you so much! I'd love to become a code owner one day and I'm really excited to contribute to this project

[ValerianRey]

LGTM.
Still need to rename typical_inputs to non-scalar inputs (per Pierre's comment).
Also I forgot to mention that we need a changelog entry (I can do it).

@PierreQuinton I think we should make a release with this.

[ppraneth]

@ValerianRey Done, I have renamed typical_inputs to non_scalar_inputs.

[ValerianRey]

Thanks! Will do the changelog + merge tomorrow.

[PierreQuinton]

Can't we use all_inputs essentially everywhere?

[ppraneth]

Done, Switched all the parametrizations across test_mean.py, test_sum.py, and test_random.py to use all_inputs. With that done, non_scalar_inputs had no remaining use, so I removed it from _inputs.py