[Bugfix] Infer the draft path for Qwen3.5 bundled MTP

[omirosh]

Motivation

Qwen3.5 already bundles its MTP draft in the target checkpoint, so --speculative-draft-model-path is redundant, but omitting it currently fails the load. This PR adds the four Qwen3.5 architectures to the existing list of models that default the draft path to --model-path.

Modifications

Add the four Qwen3.5 architectures next to Qwen4ExpForConditionalGeneration, which is listed for the same reason, and make the warning name the architecture it fired for. It read "DeepSeek MTP does not require setting speculative_draft_model_path", already wrong for the GLM, Bailing, HunYuan and Qwen4-Exp entries.

This cannot regress an existing setup. The new entries only take effect when speculative_draft_model_path is None, which previously could not work for these architectures. Callers that already pass the flag keep the path they passed.

InternS2PreviewForConditionalGeneration, InternS2MobiusForConditionalGeneration and Qwen3MoeForCausalLM have MTP remaps too and are absent for the same reason, one line each if their owners confirm the published checkpoints bundle the draft. Left out because I cannot test them.

Accuracy Tests

None needed: this resolves a launch argument to the same path users pass by hand today. Checked on amd/Qwen3.8-2.4T-A95B-Quark-MXFP4 (Qwen3_5MoeForCausalLM, NEXTN 3/1/4, TP8).

Speed Tests and Profiling

None; argument resolution only, no runtime path changes.

Related: #39064, a separate fix for mixed Quark MXFP4 Qwen3.5 MTP checkpoints.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests. _handle_eagle_family needs a resolved ModelConfig, so exercising it wants a real checkpoint config rather than a CPU unit test, and the change is four string literals. Happy to add one if you can point me at the preferred fixture.
• Update documentation according to Write documentations. Nothing to document: this removes a required flag, it does not add one.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed. Not applicable, as noted above.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ❌ Run #37437644012
Latest PR Test (Extra): ❌ Run #37437643739
Latest PR Test (AMD ROCm 10): ❌ Run #37437644040

[omirosh]

@hnyls2002 @Qiaolin-Yu could you review this? It is argument resolution only: four Qwen3.5 architectures now default --speculative-draft-model-path to --model-path when the flag is omitted, same as Qwen4-Exp. Callers who already pass the flag are unchanged. Happy to add a unit test if you have a preferred fixture.