Add Qwen3.5 model support via opt-in dependency extra

[ricky-chaoju]

Summary

Enable Qwen3.5 (dense and MoE) model support by adding a [qwen35] optional dependency extra. Base dependencies unchanged — existing users unaffected.

Consolidates and supersedes #121, #123, #129 .

Why the original PRs' runtime fixes are unnecessary

Original PR fix	Why not needed
#121 hybrid cache batched decode fallback	mlx-lm ≥0.31.0 Qwen3.5 attention correctly handles BatchKVCache.offset as mx.array — batched decode works without fallback
#123 rope validation monkeypatch	Qwen3.5 config has partial_rotary_factor=0.25, which triggers transformers' built-in set() coercion path
#129 mlx-lm model alias shim	mlx-lm ≥0.31.0 natively includes qwen3_5 and qwen3_5_moe modules

Installation

# Qwen3.5 users                                                                                                                                                         
VLLM_VERSION=0.17.0 ./install.sh                                                                                                                                        
pip install 'vllm-metal[qwen35]' 

[LxYuan0420]

Thanks for consolidating those PR. The compatibility analysis is useful and we'll reference it.

Closing this PR because:

1. transformers>=5.0.0 in an optional extra isn't isolated — it upgrades the entire environment and breaks vLLM 0.14.x + all existing models. Transformers v5 may be fine but we need to verify carefully.
2. vLLM upgrade is a project-wide migration, not a per-model opt-in. We're waiting on #134 (@ericcurtin's wheel install) to confirm the macOS path before bumping further.
3. Qwen3.5 has ongoing MLX incompatibilities — cache is broken for hybrid architectures (mlx-lm#980, mlx-lm#903), server tool calls fail (mlx-lm#905). Not ready to claim support until upstream stabilizes.
4. VLLM_VERSION env var is fragile — default install must always produce a working setup.
5. No smoke tests for existing models or end-to-end tests for Qwen3.5.

Suggested approach: Break this into steps: (1) first land the baseline upgrade with smoke tests proving existing models still work, then (2) add Qwen3.5 support once upstream MLX issuesare resolved.

Happy to review a re-scoped version. 🙏