fix: qwen3.6 moe mtp fused expert layout

[LiuXTao]

Problem

When loading Qwen3.6 MoE checkpoints with MTP (Multi-Token Prediction) layers, the bridge raises a KeyError because Qwen3.6 stores MTP expert weights in a fused/stacked layout (mtp.layers.0.mlp.experts.gate_up_proj with shape [num_experts, ...]), while the existing code only handles the Qwen3.5 per-expert layout (mtp.layers.0.mlp.experts.{i}.gate_proj.weight).

qwen3.6-35B-A3B

qwen3.5-35B-A3B

Solution

Auto-detect the MTP expert layout at checkpoint load time by checking for the fused key mtp.layers.0.mlp.experts.gate_up_proj in the safetensors index. When the fused layout is detected, map MTP expert parameters to the single stacked HF key and reuse the existing decoder MoE slice/stack logic for both import (slice by expert index from the fused tensor) and export (collect per-expert weights then torch.stack). The Qwen3.5 per-expert MTP path remains unchanged, ensuring backward compatibility.