[Qwen3.5] Add moe weight sync script for 35b model

[Rohan-Bierneni]

Description

Enables weight sync between maxtext & vllm for rl for Qwen3.5-35B.

If the change fixes a bug or a Github issue, please include a link, e.g.,:
FIXES: b/123456

Tests

Command to run validator script: https://paste.googleplex.com/5425192473067520#l=24
Result: https://paste.googleplex.com/5285887440191488

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[codecov]

Codecov Report

❌ Patch coverage is 6.58683% with 156 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...t/integration/vllm/torchax_converter/qwen35_moe.py	6.28%	149 Missing ⚠️
...ation/vllm/torchax_converter/validate_converter.py	0.00%	5 Missing ⚠️
...c/maxtext/integration/vllm/maxtext_vllm_rollout.py	33.33%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[github-actions]

🤖 Hi @aireenmei, I've received your request, and I'm working on it now! You can track my progress in the logs for more details.

[github-actions]

## 📋 Review Summary

This Pull Request successfully introduces the Qwen35MaxTextToVLLMConverter, enabling weight synchronization for the Qwen 3.5 35B MoE model between MaxText and vLLM. The implementation correctly handles the hybrid architecture (GDN + Attention) and includes necessary TPU-specific optimizations like 128-byte alignment for MoE experts.

🔍 General Feedback

• The converter logic is well-structured and follows the established patterns in the torchax_converter module.
• Proactive memory management using gc.collect() in conversion loops is a good practice for handling large model weights.
• The use of dynamic hyperparameter extraction via getattr provides good flexibility for different model variants.
• The addition of TPU-specific GMM alignment (padding to 128) ensures compatibility with vLLM's optimized TPU kernels.

[khatwanimohit]

LGTM