UnifiedTransformer -> nnsight0.5

[AdamBelfki3]

Reintroduce the implementation of the UnifiedTransformer class to work with the latest version nnsight. Add unit tests to validate this functionality.

[aaronmueller]

Any updates here? We're using UnifiedTransformer in one of our baseline implementations for MIB, and it would be super helpful to have in a more recent version!

[davidbau]

There's a discussion Jaden wants to have about whether an 0.5 feature (operation hooking) makes it unneeded to have UnifiedTransformer in this version.

[Butanium]

Yeah I think it might makes more sense to implement a new class using renaming e.g. like this and predefine hook functions per model to grab attention etc rather than using TL as a backend.

Pro:

• Official impl -> no precision issue

Cons:

• If HF change the name of intermediate variable this would break while TL impl would remain more stable