E3: qwen3_5_moe (linear attention + MoE FFN)

[tlkahn]

qwen3_5_moe (MODEL_QWEN3_5_MOE)

Linear-attention layers + MoE FFN. Separate enum from dense MODEL_QWEN3_5 (decision 4).

Spec: doc/plans/engine-inference-core.md § E3

Part of: #106 (E3) · Epic: #56
Depends on: E1 MoE block (#104); hybrid cache + unified gate

Tasks

• Linear-attention layers: separate QKV/Z/A/B projections per LinearAttnWeights
• MoE FFN via shared fwd_moe
• Layer-type mixing / config consumption already partially present from dense path
• Weight mapping + weights_expected_names() for MODEL_QWEN3_5_MOE
• Support gate allow-list entry
• Smoke generate on canonical MoE checkpoint

Notes

• Do not collapse with MODEL_QWEN3_5 dense