[Feature Request] Add support for Qwen3.5/3.6 MoE (qwen3_5_moe) architecture

[shrey2003]

Is your feature request related to a problem? Please describe.

Attempting to export Qwen/Qwen3.6-35B-A3B (and other Qwen3.5/3.6 MoE variants) using QEfficient.cloud.export fails at the config loading stage because qwen3_5_moe is not a recognized architecture in the transformers version that QEfficient currently pins:

KeyError: 'qwen3_5_moe'
ValueError: The checkpoint you are trying to load has model type `qwen3_5_moe`
but Transformers does not recognize this architecture.

The qwen3_5_moe architecture was introduced in transformers>=5.3.0, but QEfficient's current dependency tree pins transformers<5.x, making it impossible to export or compile Qwen3.5/3.6 MoE models for AIC100 without breaking QEfficient's own internal imports.

Describe the solution you'd like

1. Add qwen3_5_moe / Qwen3_5MoeForCausalLM to QEfficient's supported model registry (QEfficient/utils/model_registery.py)
2. Add the necessary AIC100-specific PyTorch transforms for the Qwen3.5 MoE attention and MoE routing layers in QEfficient/transformers/models/
3. Bump the transformers dependency pin to >=5.3.0 (while fixing the internal import breakages caused by symbols removed in transformers>=5.4.0, such as AwqBackendPackingMethod, AWQLinearVersion, HybridCache, and Qwen2RMSNorm from qwen2_5_vl)
4. Add an ONNX export config for text-generation-with-past task for this architecture

Describe alternatives you've considered

• Manually patching the installed QEfficient venv files with sed to stub out missing imports — this is unsustainable as each fix exposes another broken import down the chain
• Using optimum-cli export onnx directly — same blocker, as optimum also has a transformers>=5.x compatibility issue
• Downgrading to a supported Qwen variant (e.g., Qwen2-57B-A14B) — undesirable as Qwen3.5/3.6 MoE offers significantly better performance per parameter

Additional context

• Target model: Qwen/Qwen3.6-35B-A3B (35B total params, 3.6B active, qwen3_5_moe architecture)
• Hardware: Qualcomm AIC100 (ai-ABPI-130, Ubuntu 22.04)
• QEfficient version: installed from PyPI (latest), Python 3.10
• transformers version that supports the arch: 5.3.0 (works for config loading; >=5.4.0 has a regression — see [huggingface/transformers#45310]([BUG] transformers>=5.4.0, Qwen3.5 Moe from_pretrained error huggingface/transformers#45310))
• The Qwen3.5 MoE family (35B-A3B, 30B-A3B, 122B-A14B) is increasingly popular for on-device and edge inference, making AIC100 support particularly valuable

---