[OpenVINO] Add Support for GroupedMM in AWQ and Scale Estimation for OpenVINO Backend

[anzr299]

Changes

GroupedMatMul applies one weight matrix per group to a single shared 2D activation. Data-aware compression assumed a 3D weight always comes with a matching 3D activation, so AWQ, scale estimation and GPTQ failed.

AWQ

For weights, The experts axis is folded into the output channels, the search then minimizes the error over all experts at once, and the scale is broadcast back over the experts axis.

For activation, the same scale is applied over the hidden dim of the activation in non mergible case, For mergible scale, it is reshaped but still only applied over the hidden dim of the previous matmul weights.

Scale estimation

Broadcasts the pooled statistics over the group axis, so the scale stays per group.

Reason for changes

to support grouped_mm based MoE models in OpenVINO backend.

Related tickets

191895

Tests

test_scale_estimation and test_awq_scale_reference extended for this case.
test_node_utils extended for the weight channel axes and activation channel axis for GroupedMatMul.

WWB Eval

Model: Qwen3.6 35B A3B
mode: INT4_ASYM; GS: 64;

AWQ=True

MoE op	WWB Accuracy
BatchedMM	0.956345
GroupedMM	0.952167

Scale Estimation=True

MoE op	WWB Accuracy
BatchedMM	0.9534
GroupedMM	0.9587

Test examples - success

Weight compression - success

[andreyanufr]

Is it possible to move reference to json file ? We have examples in tests.

[anzr299]

That is a good idea. I will move it.

[anzr299]

I have combined all the refs for SE in 1 function (earlier MoE refs had another getter). Then moved it to a reference JSON file.