refactor(aggregation)!: Remove generalized gramians

[ValerianRey]

Closes #690

• Engine.compute_gramian now always returns a flat [m, m] gramian, regardless of the output shape
• Removed GeneralizedWeighting, and Flattening — they are no longer needed
• Updated the IWMTL example to use UPGradWeighting directly and reshape the weights before calling backward
• Added a migration guide in CHANGELOG.md

[PierreQuinton]

I think I like it but let me think about it. Also I don't think we want to erase all private generalized_gramian utilities and put their implementation in autogram.

[ValerianRey]

@PierreQuinton this is now ready for review. I reverted to (almost) the original implementation of autogram: compute square gramian, reshape, movedim. The extra flatten is because we now want to flatten.
I also reviewed everything.

In a future PR we should be able to let compute_gramian take any number of output tensors (instead of a single one), to match the interface of autojac. There are a bunch of other changes to make the interfaces of autogram and autojac closer together, but let's go step by step.

[PierreQuinton]

In a future PR we should be able to let compute_gramian take any number of output tensors (instead of a single one), to match the interface of autojac. There are a bunch of other changes to make the interfaces of autogram and autojac closer together, but let's go step by step.

That would be nice. But I think that this is not trivial with the batched dims (I guess all could have one, and possibly a different? I think in principle that would work but not trivial).