[GPU] Fix FullyConnected accuracy regression for large f16 matmuls on non-immad devices

[alvoron]

Details:

• [GPU] Selectively keep FullyConnected weights non-transposed for large matmuls #36437 introduced a weights_transposed (transpose_b) flag and, for large non-compressed f16 matmuls (K >= 8192 && (M <= 512 || N <= 4096)), keeps the weights in the non-transposed [K, N] layout so they can be executed with the faster onednn abc weight layout.
• However, the non-transposed weight layout is only consumed by the onednn FullyConnected implementation. The OCL FullyConnected kernels always assume transposed [N, K] weights and ignore the flag - they only reference weights_transposed for output-shape calculation, not for weight indexing.
• On devices without immad support, the OCL FC path is used. There, feeding non-transposed weights while the kernel indexes them as transposed makes the kernel read the wrong weights.

Fix:

• The fix gates the non-transposed-weights decision on supports_immad, so the layout is only chosen when onednn will actually consume it. On non-immad devices the pass now keeps the previous transposed behavior.

Tickets:

• CVS-190678

[maxnick]

@clee30 , could you please review?

[Copilot]

Pull request overview

Fixes an Intel GPU plugin accuracy regression where large f16 MatMul→FullyConnected conversions could keep weights in non-transposed [K, N] layout even when the execution path uses OCL FullyConnected kernels (non-immad devices) that assume transposed weights, causing incorrect weight indexing.

Changes:

• Gate the “keep weights non-transposed for large f16 matmuls” heuristic on supports_immad, ensuring the non-transposed layout is only selected when the oneDNN FullyConnected implementation will consume it.
• Add a unit test covering the non-immad case to ensure the transformation keeps the transposed behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/plugins/intel_gpu/src/plugin/transformations/convert_matmul_to_fc.cpp	Gates the large-f16 non-transposed weights decision on supports_immad to avoid OCL FC misinterpreting weight layout.
src/plugins/intel_gpu/tests/unit/transformations/convert_matmul_to_fc_test.cpp	Adds a regression test ensuring non-immad devices keep the transposed weights path.

[alvoron]

@maxnick could you please review?