feat(autojac): Make jac_to_grad return optional weights

[ValerianRey]

This is based on #585 but instead of taking a method argument that can be either Aggregator or Weighting, it always takes an aggregator.

• feat(autojac): Make jac_to_grad return optional weights
• Add test with Weighting
• Make jac_to_grad only take aggregator
• Make the aggregator parameter positional & keyword

There's just one problem: currently, WeightedAggregator is not public. So we can't really talk about this class in the docstring of jac_to_grad. I think we have two options:

• Make WeightedAggregator public. Because it has a weighting parameter, of type Weighting[Matrix], we'll have to find a solution for that too (make Matrix public or annotate weighting as Weighting).
• Say something like "If the aggregator is based on a Weighting" instead of "If a :class:WeightedAggregator <torchjd.aggregation._aggregator_bases.WeightedAggregator> is provided"

A similar issue will arise when we'll add the option to do the gramian-sum optimization. We'll have to talk about GramianWeightedAggregator and make it public (as well as PSDMatrix), or we'll have to say "aggregator uses a gramian-based Weighting`.

Note that this latter problem would also happen to some extent in #585, because Weighting[PSDMatrix] is not public, so we'll also have to say something like if weighting is gramian-based or make PSDMatrix public.

I think the easiest solution for now is just to talk about "aggregator using a Weighting", and later about "aggregator using a gramian-based Weighting" when we add the gramian-sum optimization. (Done in b701c59)