fix(qwen4-exp): keep text inference on mlx-vlm path

[Thump604]

Qwen4-Exp is not compatible with the generic Qwen3.5 TextModel fallback. Keep text requests on the loaded mlx-vlm language model instead of constructing a mechanically compatible model that produces incorrect logits.

Checks:

• 13 focused text-model dispatch tests
• 954 non-integration Apple-Silicon tests
• Ruff and Black

[janhilgard]

Correct fix for the right reason. Returning None so SimpleEngine stays on the loaded mlx-vlm model is better than the alternative here: a mechanically compatible skeleton that loads cleanly and emits wrong logits is far harder to diagnose than a model that refuses the fast path, because nothing fails — the output is just quietly worse.

Two details I checked rather than assumed:

• str.startswith accepts a tuple, so _VLM_ONLY_TEXT_MODEL_PREFIXES works as written and covers qwen4_exp_text from the single entry — the parametrised test pins both spellings, which is what makes the prefix form safe to extend later
• the check sits before _import_text_model_classes, so an unregistered architecture short-circuits instead of falling through to the qwen3_5 default that #686 made prefix-based

The logger.info naming the model type matters more than it looks: this path silently changes which engine serves text, and without that line the only symptom would be a throughput difference nobody attributes to dispatch.

Approving.

[waybarrios]

all set and ready to go