Fix re-compilations for cross attention cache

[zucchini-nlp]

What does this PR do?

Fixes #39774.

As per title, if we are using the legacy cache.key_cache[layer_idx] a warning is emitted and fullgraph compilation breaks. This PR makes sure no warning are raised when using the models in core library

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: autoformer, bert, bert_generation, big_bird, bigbird_pegasus, blip, bridgetower, camembert, data2vec, electra, ernie, fsmt, gpt_bigcode, imagegpt, kosmos2, led

[manueldeprada]

lgtm, sorry!! these changes got lost when cherry-picking back and forth between the layer[i].keys and key_cache[i] designs in the original PR😭

[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.

[zucchini-nlp]

No worries, that happens 😄

Let me see if I can add encoder-decoder compile test easily in this PR or if we need to handle a lot of edge cases

EDIT: oh these aren't generative models/can't compile fullgraph and we don't have graph-break test for those models yet. That's why it wasn't caught in CI