Ensure same dtype for subconfig when _from_config

[zucchini-nlp]

What does this PR do?

Fixes "auto" dtype when the model is initialized from_config

It was already fixed for from_pretrained in #42990 but vLLM creates models with AutoModel._from_config which caused the same dtype mismatch error

Added a test as well. cc @hmellor

---

Note

Medium Risk
Touches core model initialization and could affect dtype override behavior for any model with sub_configs, though the change is narrow (dtype propagation) and covered by a targeted regression test.

Overview
Fixes dtype mismatches when instantiating nested/compound models via PreTrainedModel._from_config by forcing the chosen dtype onto all config.sub_configs (mirroring from_pretrained behavior), ensuring modules don’t initialize with conflicting dtypes (notably for dtype="auto" paths used by vLLM/AutoModel).

Adds a regression test around LlavaForConditionalGeneration._from_config verifying that a parent config.dtype overrides differing sub-config dtypes and results in a consistent dtype across language/vision components and the top-level model.

^{Written by Cursor Bugbot for commit 5a895f4. This will update automatically on new commits. Configure here.}

[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.

[hmellor]

Will this ever be None? The type hint suggests that it shouldn't be

[zucchini-nlp]

the subconfig? I remember it was possible in some backbone_configs in pure-vision models. In generative mllm it doesn't happen

Actually I started refactoring those backbones so maybe not needed anymore, then I can clean up everywhere

[hmellor]

Yeah the values in config.sub_configs. I was just curious because it's hinted as dict[str, type["PreTrainedConfig"]]. Checking for None doesn't hurt though