Decorate cache updates with no_grad, just in case

[Rocketknight1]

Although our cache update methods are usually used in inference, when grad is disabled anyway, there seem to be some edge cases where they cause problems with compilation and gradient computation. Since we never want to propagate gradient through these steps, adding @no_grad decorators shouldn't hurt, and fixes the edge cases.

Fixes #43010

[github-actions]

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

run-slow: bamba, falcon_h1, falcon_mamba, granitemoehybrid, jamba, mamba, mamba2

[Rocketknight1]

Not sure who the right person to ping for generate/caches is now so cc @remi-or @ArthurZucker @McPatate

[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.

[ArthurZucker]

ty

[Cyrilvallez]

Once again the test fetcher bites us... This fails a few test_retain_grad_hidden_states_attentions and test_training_ci
cc @ArthurZucker @ydshieh @tarekziade