[CPU] Introduce GatherMatmul operation to optimize MoE pattern

[maxnick]

Details:

In this PR we introduce yet another operation "GatherMatmu", which essentially does gemv operations over the current tokens and the active experts.
As the first step, we perform gemv operation using the dnnl::inner_product. But obviously this solution is suboptimal, as it doesn't give a fine grain control over parallelization, and in the case of many tokens being processed by a specific expert (prefill), having gemm operation may be more optimal as the tokens may be batched and we can do SIMD level parallelization by tokens as well.
Also this PR contains all the essential transformations that allow to enable a few common MoE patterns.

MoE pattern matcher is based on #32183

Related oneDNN fork PR: openvinotoolkit/oneDNN#292

Tickets:

• CVS-171910

[Copilot]

Pull Request Overview

This PR introduces the GatherMatmul operation to optimize Mixture of Experts (MoE) patterns in the CPU plugin. The implementation performs GEMV operations over active experts using oneDNN's inner_product primitive, with support for both standard and compressed weights configurations.

Key changes:

• Adds GatherMatmul node implementation with oneDNN-based execution for both GEMV and GEMM modes
• Implements pattern matchers (MoE2GeMM and MoE3GeMM) to detect and transform MoE subgraphs
• Extends weight decompression infrastructure to support batched (3D) weight tensors
• Introduces CompressedWeightsBlock pattern block to share weight decompression logic across operations

Reviewed Changes

Copilot reviewed 31 out of 31 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/tests/functional/plugin/shared/src/subgraph/weights_decompression_builders.cpp	Updated to support batched weight decompression with seed parameter and optional transpose control
src/tests/functional/plugin/shared/src/subgraph/moe_builders.cpp	New MoE test graph builders for 2GEMM and 3GEMM patterns with weight decompression support
src/plugins/intel_cpu/src/nodes/gathermatmul.cpp	Core implementation of GatherMatmul node with oneDNN inner_product backend
src/plugins/intel_cpu/src/transformations/cpu_opset/common/pass/convert_moe_matmuls.cpp	Pattern matchers to detect and replace MoE patterns with BatchGatherMatmul operations
src/plugins/intel_cpu/src/transformations/cpu_opset/common/op/batch_gather_matmul*.cpp	New internal operations for batch gather matmul with and without compression
src/common/transformations/src/transformations/op_conversions/convert_fc_to_compressed.cpp	Refactored to extract reusable weight processing logic into static method
src/core/include/openvino/pass/pattern/op/block_util.hpp	Updated FOR_EACH macros to support passing block pointer as parameter

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.