Fix MXFP4 expert zero weight loading

[robbiemu]

Summary

• distinguish the all-expert None sentinel from valid expert ID 0 in the static MXFP4 weight loader
• add a focused CPU regression for a 2-D per-expert MXFP4 weight with expert ID 0

Root cause

FusedMoE.weight_loader used not expert_id to detect an all-expert tensor. That condition is also true for expert ID 0, so the loader entered the 3-D all-expert path and indexed loaded_weight.shape[2] on a 2-D per-expert tensor.

This small fix is also present in the broader open #21529; this PR splits the loader correctness change and its regression test into an independently reviewable patch.

Impact

Static MXFP4 checkpoints loaded one expert at a time can fail while loading expert zero with IndexError: tuple index out of range.

Validation

• python3 -m pytest -q test/registered/unit/layers/moe/test_mxfp4_expert_weight_loader.py
• python3 -m ruff check python/sglang/srt/layers/moe/fused_moe_triton/layer.py test/registered/unit/layers/moe/test_mxfp4_expert_weight_loader.py

---

CI States

Latest PR Test (Base): ❌ Run #29341502637
Latest PR Test (Extra): ❌ Run #29341502576

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[robbiemu]

This probably should have been merged before #32541 . You can feel free to close it.

Maybe there is a norm I'm not conforming to that my PRs get no review and the same code gets merged later? (Of course, I understand things slip through the cracks sometimes! Just want to make sure I understand if and when contributions are accepted)