Harmonize `past_key_value` to `past_key_valueS` everywhere (#39956)
* all modulars and llama
* apply modular
* bert and gpt2 copies
* fix imports
* do it everywhere
* fix import
* finalize it
* fix
* oups set it in modular
* style
* fix
* Add 1 version to deprecation cycle
* Update modeling_layers.py