[Feature] Support Qwen3.5 text model

[botieking98]

Background

This PR adds Qwen3.5 text-only model support to mini-sglang, while keeping the serving path aligned with existing mini-sglang/sglang runtime patterns.

What Changed

1. Added Qwen3.5 model implementation and model registration.
2. Added the required GDN/Radix linear-attention path for Qwen3.5 prefill/decode.
3. Introduced minimal runtime infrastructure needed by this path:

• forward-batch context wiring
• custom op registration helpers
• model-executor helpers
• compilation context utilities

4. Updated related attention/layer/scheduler/tokenizer/model-config wiring so /v1/chat/completions works end-to-end.

Scope

• In scope: text inference path only for Qwen3.5 family.
• Out of scope: multimodal support, broader refactors unrelated to Qwen3.5.

Validation

python -m minisgl --model /root/models/Qwen3.6-27B --tp 8 --port 30000 --dtype bfloat16 --attn fi

Notes for Reviewers

• This PR touches multiple subsystems because Qwen3.5 depends on both model-side and attention backend/kernel-side support.
• Interfaces were kept as close as practical to upstream sglang patterns to reduce future sync cost.