Qwen3.6 (qwen3_5 / qwen3_5_moe) text-generation export blocked — only image-text-to-text task registered

Nobody has claimed this yet.

## Summary

Both `Qwen/Qwen3.6-35B-A3B` (MoE) and `Qwen/Qwen3.6-27B` (dense) are published by the Qwen team as unified VL variants with `architectures: ['Qwen3_5MoeForConditionalGeneration']` / `['Qwen3_5ForConditionalGeneration']`. There are no separate text-only releases.

The OpenVINO exporter on `main` (`1.27.0.dev0+d4dd21a` as of 2026-05-05) registers these parent architectures for `image-text-to-text` only, so attempting a text-generation export aborts:

```
$ optimum-cli export openvino \
--model Qwen/Qwen3.6-35B-A3B \
--task text-generation-with-past \
--weight-format int4 \
--group-size 64 \
--ratio 1.0 \
out/

ValueError: Asked to export a qwen3_5_moe model for the task text-generation-with-past,
but the Optimum OpenVINO exporter only supports the tasks image-text-to-text for qwen3_5_moe.
Please use a supported task. Please open an issue at https://github.com/huggingface/optimum-intel/issues
if you would like the task text-generation-with-past to be supported in the OpenVINO export for qwen3_5_moe.
```

Same error for `Qwen/Qwen3.6-27B` (`qwen3_5`).

## Repro

```bash
pip install \
"optimum-intel @ git+https://github.com/huggingface/optimum-intel.git@main" \
"transformers @ git+https://github.com/huggingface/transformers.git@main"

optimum-cli export openvino \
--model Qwen/Qwen3.6-27B \
--task text-generation-with-past \
--weight-format int4 \
out/
```

(transformers main is required separately — release < 5.7 doesn't recognize the `qwen3_5` / `qwen3_5_moe` model_types.)

## Root cause

`optimum/exporters/openvino/model_configs.py` has the registrations split:

- `qwen3_5_moe` (line ~6174) → `image-text-to-text` only — corresponds to the unified `*_ForConditionalGeneration` architecture HF publishes
- `qwen3_5_moe_text` (line ~6165) → `text-generation`, `text-generation-with-past` — works but only routable from inside the VL config's LANGUAGE behavior code path, not as an export entrypoint
- Same split for `qwen3_5` / `qwen3_5_text`

The text-only sub-configs do work: `Qwen3_5MoeOpenVINOConfig` already delegates to `qwen3_5_moe_text` via `get_vlm_text_generation_config` for the `LANGUAGE` behavior. So the conversion machinery is in place — only the entry-point routing is missing.

## Suggested fix

Add `text-generation` / `text-generation-with-past` to the `qwen3_5_moe` and `qwen3_5` task registrations. When the user requests text-generation, the exporter should route to `LANGUAGE` behavior (extract `text_config` from the unified config, run the text-only path).

Pattern would mirror the `gemma3` / `gemma-4` family which has `image-text-to-text` and `text-generation` both registered against the same parent architecture.

## Why it matters

These are widely-used Qwen models (~3M and ~1.5M HF downloads). Without the text-generation export route, OpenVINO users have to either:
- Custom-extract `model.language_model` and re-save as a text-only Qwen3_5MoeForCausalLM (several hours of careful work)
- Patch optimum-intel locally (fragile across pip upgrades)
- Wait

If a quick fix is welcome, I can put together a PR. Just wanted to log the issue so others hitting this aren't stuck guessing.

Start in optimum/exporters/openvino/model_configs.py around the qwen3_5 and qwen3_5_moe registrations, then compare the task routing used by the gemma3 and gemma-4 families. Add the text-generation task entry points so unified Qwen3.6 configurations use their LANGUAGE behavior, and verify that the provided optimum-cli text-generation-with-past export commands complete successfully.

Written by the indexing model from the issue text.