Add Qwen3.5 support for sequence classification

[medhakimbedhief]

Adds sequence-classification support for Qwen3.5 in AutoModelForSequenceClassification.

What does this PR do?
This PR enables loading Qwen3.5 checkpoints with AutoModelForSequenceClassification, which previously failed with:
ValueError: Unrecognized configuration class Qwen3_5Config for AutoModelForSequenceClassification.

Changes

• Adds Qwen3_5ForSequenceClassification in:
  -- src/transformers/models/qwen3_5/modular_qwen3_5.py
  -- generated src/transformers/models/qwen3_5/modeling_qwen3_5.py
• Exports the new class via the Qwen3.5 modeling __all__.
• Registers auto-mapping entries in:
  -- src/transformers/models/auto/modeling_auto.py
  -- ("qwen3_5", "Qwen3_5ForSequenceClassification")
  -- ("qwen3_5_text", "Qwen3_5ForSequenceClassification")

Why both mappings?
Qwen3.5 uses a composite VLM config (qwen3_5) with a text sub-config (qwen3_5_text).
Registering both ensures classification works for direct text config usage and composite config loading paths.

Before :
Loading from e.g. Qwen/Qwen3.5-0.8B raises ValueError: Unrecognized configuration class Qwen3_5Config for AutoModelForSequenceClassification.

After this PR:
Loading from Qwen/Qwen3.5-0.8B now resolves to Qwen3_5ForSequenceClassification.

@Cyrilvallez @zucchini-nlp @ArthurZucker

Fixes #44405

[github-actions]

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

run-slow: auto, qwen3_5

[zucchini-nlp]

The PR lgtm. We usually don't add a task class for models if there are no weights on the hub for this specific task. Though with the GenericSequenceClassfier I think we can even automatically create sequence classification tasks on top on base models

[medhakimbedhief]

Just to confirm @zucchini-nlp, are you suggesting we enable sequence classification generically for base checkpoints like Qwen3.5-0.8B-Base via GenericForSequenceClassification, rather than adding a dedicated Qwen3_5ForSequenceClassification class?

That makes sense regarding not expanding the task surface when no task-specific weights exist on the Hub.

[zucchini-nlp]

Ah no, it's fine for this PR. I was just thinking out loud :)

Let's merge then

[mnbucher]

I saw that the current sequence classifier is configured via

config: Qwen3_5TextConfig

https://github.com/huggingface/transformers/pull/44406/changes#diff-403c772331fa104c97f12712903c05e01f6335afdfa1f1719a52ed6ea7d1206dR1875

which seems to ignore the vision tower and is not ideal for VLM-type tasks. is this intended? i'm not very familiar with the whole GenericForSequenceClassification idea etc...

[zucchini-nlp]

nope, kinda related to #44625. Sorry, it got swept off as a lower prio task from my list, will do smth with it