[configuration] allow to overwrite kwargs from subconfigs

[zucchini-nlp]

What does this PR do?

Fixes #38258 and allows users to overwrite kwargs from subconfigs in the same way as we allow to overwrite in the main config. See below code for what has changed

# Create a config that is same as the LLaVa-NeXT Mistral with smaller vocab size

# Before
config = AutoConfig.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf", text_config={"vocab_size": 900})
config.text_config.sliding_window 
>>> 4096 # default value for sliding_window in Mistral config
config.text_config.vocab_size
>>> 900  # Vocab size was changed to user-defined kwarg

# After
config = AutoConfig.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf", text_config={"vocab_size": 900})
config.text_config.sliding_window
>>> None # sliding_window from loaded checkpoint
config.text_config.vocab_size
>>> 900  # Vocab size was changed to user-defined kwarg

[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]

Ish a bit scared by it but should be alright, let's run some slow tests on the llava / qwenvl models

[zucchini-nlp]

run-slow: llava, qwen2_vl, gemma3, qwen2_5_vl

[github-actions]

This comment contains run-slow, running the specified jobs:

models: ['models/gemma3', 'models/llava', 'models/qwen2_5_vl', 'models/qwen2_vl']
quantizations: [] ...

[zucchini-nlp]

Failing slow tests are same as in main, so we're fine