[MLAS] Removed memcpy step by storing result in C if possible

[JonathanC-ARM]

Summary

This change removes the memcpy step in sgemm_kleidiai where possible by writing directly to C

Testing

Model	Baseline avg (ms)	Current avg (ms)	Δ ms	Δ %
Transformer_complex_f32.onnx	2.929885	2.701083	-0.228802	-7.81%
bert_tiny_f32.onnx	0.279675	0.273928	-0.005747	-2.05%
de_efficientnetlitev3_f32.onnx	80.038132	78.560747	-1.477385	-1.85%
deeplabv3_mobilenetv2_f32.onnx	48.565125	46.446841	-2.118284	-4.36%
imagetransformnet_f32.onnx	303.835868	302.553625	-1.282243	-0.42%
mobilenet_v1_f32.onnx	4.379468	4.163018	-0.216450	-4.94%
mobilenetv1_ssd_f32.onnx	9.245055	8.881198	-0.363857	-3.94%
openposev2_vgg19_f32.onnx	210.981128	209.199398	-1.781730	-0.84%
retinaface_f32.onnx	42.326391	38.454346	-3.872045	-9.15%
rfdn_f32.onnx	13.929565	13.679875	-0.249690	-1.79%

[hariharans29]

/azp run Linux QNN CI Pipeline,Win_TRT_Minimal_CUDA_Test_CI,Windows ARM64 QNN CI Pipeline,Windows GPU Doc Gen CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 4 pipeline(s).

[Copilot]

Pull request overview

This PR optimizes MLAS KleidiAI's SGEMM implementation by eliminating an unnecessary memcpy step. When alpha=1.0 and beta=0.0, the result is now written directly to the output matrix C instead of first writing to a temporary buffer and then copying.

Changes:

• Simplified the direct-write condition from checking multiple constraints (ldc==TileSizeN, boundary checks, zero size checks) to only checking alpha==1.0 and beta==0.0
• Leverages KleidiAI's run_matmul capability to write directly to non-contiguous output via the row stride parameter
• Eliminates temporary buffer allocation and memcpy when alpha/beta scaling is not needed

---

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