🚨🚨 Fix and simplify attention implementation dispatch and subconfigs handling (#39423)

* first try

* Update modeling_utils.py

* Update modeling_utils.py

* big refactor

* Update modeling_utils.py

* style

* docstrings and simplify inner workings of configs

* remove all trace of _internal

* Update modeling_utils.py

* fix logic error

* Update modeling_utils.py

* recursive on config

* Update configuration_utils.py

* fix

* Update configuration_dpt.py

* Update configuration_utils.py

* Update configuration_utils.py

* Update modeling_idefics.py

* Update modeling_utils.py

* fix for old models

* more old models fixup

* Update modeling_utils.py

* Update configuration_utils.py

* Remove outdated test

* remove the deepcopy!! 🥵🥵

* Update test_modeling_gpt_bigcode.py

* fix qwen dispatch

* restrict to only models supporting it

* style

* switch name

* Update modeling_utils.py

* Update modeling_utils.py

* add tests!

* fix

* rypo

* remove bad copies

* fix

* Update modeling_utils.py

* additional check

* Update modeling_utils.py

* Update modeling_utils.py

* Update modeling_utils.py

* Update modeling_utils.py

* Update modeling_utils.py

* fix

* skip

docs/source/en/attention_interface.md

@@ -60,11 +60,11 @@ You will see it prints "I just entered the attention computation" as many times
 
 ## Dynamically switching attention function
 
-You could dynamically change the model's attention function as well, by overriding the `config._attn_implementation` field:
+You could dynamically change the model's attention function as well:
 
 ```python
 # Back to use original sdpa implementation
-model.config._attn_implementation = "sdpa"
+model.set_attn_implementation("sdpa")
 
 model(torch.ones(1, 5, dtype=int))
 ```