Deal with weight tying in transformers >=5

[githubnemo]

While we already implemented forward compatibility with the way transformers>=5 handles weight tying, there was an issue with weight tying of trainable tokens wrappers.

Before, we simply got fixed strings of which modules are tied to the embeddings, e.g. "lm_head" - this never changed since it was just a static property of the respective PretrainedModel class. However, with the new way get_tied_weights_keys is implemented, the names of the tied-to-embeddings modules change if they are moved around. So if we wrap the lm_head once in a trainable tokens wrapper, it'll become lm_head.token_adapter.base_layer instead of lm_head. That means that the check to see if we already wrapped the tied layer needs to look at the grand-parent instead of the target layer.

This obviously assumes that we always have a nesting level of two which is true for TrainableTokensWrapper.

[githubnemo]

I think ModulesToSaveWrapper is not relevant here since it only handles weight tying explicitly using PeftConfig.modules_to_tie.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[BenjaminBossan]

Thanks, LGTM.