qwen35 moe

[wine99]

@cavusmustafa Changes from your PR #221 included here

1. when use_bias=true, use (w - zp_f16) * scale instead of w * scale + bias_f16). For quantized expert weights, use_bias is always true
2. changes in other ops (eg process_view)

Differences from your PR:

1. use rank-3 shape for expert weights after dequantaztion subgraph (ie rank-4 before dequant) instead of rank-2. So that both CPU and GPU recognize the dequantization pattern by default and can fuse to GatherMatmulGather
2. copied GatherMatmul header file and construct it directly, like what did with GatedDeltaNet
3. for CPU, this PR [CPU][Snippets] Skip tokenizing GatherMatmul's dequantization weights path openvinotoolkit/openvino#36937 is needed
4. for CPU and GPU [CPU][GPU] Recognize GatherMatmul as a decompression-multiply consumer openvinotoolkit/openvino#36936 is needed

Both PRs in ov are needed. Without the first one, the dequantization subgraph will be converted to a snippet in ov cpu, preventing GatherMatmulCompressed fusion. Without the second one, ov::pass::KeepConstPrecision will not be called and u4->u8 conversion will happen on both CPU and GPU. Once the the second PR is merged, we can also remove the copied mark_dequantzation_subgraph.hpp

[wine99]

current status:

• test_backend_ops -o "MUL_MAT_ID" pass
• GatherMatmulCompressed op seen in CPU and GPU exec graph with a small synthetic qwen35 moe model
• llama-simple Qwen3.6-35B-A3B-UD-Q4_K_S OOM i7-14700K 64GB ubuntu throughput is 4.98 tokens/second.