Configuration insoncistencies

[zucchini-nlp]

What does this PR do?

Fixes:

• Replace PretrainedConfig with PreTrainedConfig
• Don't import from other models inside configuration_model.py, instead resolve via modular
• Text/vision sub-configs are documented (only kosmos was missing)
• Qwen-vision model type is changed, otherwise we have the same model type for Qwen2VLConfig and Qwen2VLVisionConfig. Vision backbones are not auto-mapped yet and were not even a PreTrainedModel in the past. So it won't break anything, and official checkpoint don't have model_type saved anyway
• Nougat has its own config class, instead of mapping to VisionEncoderDecoderConfig with a different model type. Rn in config's auto mapping, the VisionEncoderDecoderConfig resolves to vision-encoder-decoder by accident, we are just lucky it comes last and overwrites nougat. IMO no config class should map to two model types, since model type is a ClassVar

I wonder if we should be adding linter checks on a few of these, specifically no importing from other models and each config class maps to single model_type?

P.S. Uncovered everything when trying to dynamically build the MAPPING in #45018

[zucchini-nlp]

model_type = musicflamingo_encoder doesn't exist

[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.

[zucchini-nlp]

I think we can keep it as part of the bigger PR, closing

[github-actions]

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

run-slow: eurobert, fastspeech2_conformer, kosmos2, musicflamingo, nanochat, nougat, qwen2_5_vl, qwen2_vl, qwen3_5, qwen3_5_moe, qwen3_vl, qwen3_vl_moe, speecht5, vibevoice_asr, voxtral_realtime