feat: Upgrade to transformers v5

[vivekkalyan]

This PR upgrades transformers v5 in ART by using a fork of vLLM: https://github.com/vivekkalyan/vllm/releases/tag/v0.17.0-art1

Some patches in our ART runtime are needed to make transformersv5/Qwen3.5 work:

• Patch 1: make Transformers v5 rope config parsing not crash.
• Patch 2: make Qwen3 VL MoE config look like what vLLM expects.
• Patch 3: make Qwen3.5 fused GDN LoRA actually work in vLLM.

As well as some tokenizer shape changes to convert the newer tokenizer return shapes back into the old “token ids list” contract that the rest of preprocessing already expects.