Optimize MoE expert aggregation with ReduceSums

[wine99]

On top of the Qwen35 moe PR.

Optimizes MoE expert aggregation in the OpenVINO backend. It detects the final ADD in the structurally identified expert-view aggregation chain and translates it as a single ReduceSum over the expert dimension, instead of materializing individual expert views and chaining multiple additions.

[wine99]

Test llama-simple -m Qwen3.6-35B-A3B-UD-Q4_K_S on Raptor Lake (i7-14700K 64GB Ubuntu24)
CPU: Almost no improvements 7.28 t/s -> 7.32 t/s (ggml cpu is 12.39 t/s, vulkan 7.53 t/s)
GPU: Cannot test due to no available GatherMatmulCompressed kernel on this platform