Add Qwen3.5 (text + multimodal) support

[86MaxCao]

Summary

• Add Qwen3.5 (text + multimodal) support and loader wiring so nano-vllm can run vision–language workloads.
• Extend engine components for multimodal prefill (placeholder expansion), vision-token caching, KV / block handling for special vision tokens, and decode-time recurrent state where needed.
• Speed up GatedDeltaNet (GDN) decode with a persistent slot pool and more vectorized index-style ops, reducing heavy per-step Python gather/scatter on the hot decode path.
• Implement relevant hot paths with Triton kernels.
• Add example_qwen3_5.py and bench_qwen3_5.py for quick testing and benchmarking.
• Update README with model download notes and how to run the multimodal example.

Performance (reference)

• Text (256 seqs): ~5,338 tok/s
• Multimodal (10 images): ~550 tok/s

Testing

python3 example_qwen3_5.py --model "/YOUR/MODEL/PATH"

Notes

• This is a large diff across model loading, runner, scheduler, and caching; happy to walk through details or split the change if maintainers prefer.
• If keeping multimodal in core is not desired yet, I’m open to a slimmer PR or an extension-style layout.