refactor(aggregation): Fix typing issue in Weighting

[ValerianRey]

Since #522, we don't use type annotations anymore, so a new (user-facing) typing issue that we already had for a long time is now visible: in autogram usage examples, we call weighting(gramian) where gramian is a Tensor (but not a PSDMatrix), but weighting.__call__ expects a PSDMatrix.

I changed weighting.__call__ to rather expect a Tensor (like Aggregator.__call__ expects a Tensor and not a Matrix). I thus also made the _T type variable bound to Tensor.

This fixes the issue, but it also doesn't seem to be a very good long-term solution. I don't really like that the user-facing types "lie" to the user. Maybe we could make those types (Matrix and PSDMatrix) public at some point?

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
src/torchjd/aggregation/_weighting_bases.py	100.00% <100.00%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[PierreQuinton]

Good enough for now, but yeah, one day we might make them public though, even if this can be annoying (for isntance the function torch.eye doesn't returns a PSDMatrix, but it should). I think for now, those types are just for us to not screw things up, but the interface should be agnostic to them.