Add QDQ scale propagation pass

[javier-intel]

Description

Adding pass to propagate scale values with a magnitude above a certain threshold to avoid numerical overflows.
https://jira.devtools.intel.com/browse/CVS-170179

Motivation and Context

Improve precision on certain networks

[preetha-intel]

Is this pass happening even for non quantized models?

[mklimenk]

@preetha-intel, this pass is happening only when the enable_qdq_optimizer flag is set.
Inside the pass it specifically looks for quantized blocks with (u)int16 precision and ignores everything else. So the regular models are not affected by it, even if the flag was passed by accident

[harihs1729]

Accuracy results from PSD model testing on GPU align with both NPU outputs and Microsoft’s expectations. Please proceed with merging this PR.

[Copilot]

Pull Request Overview

Adds a new pass to propagate and adjust quantization scales in QDQ (Quantize–Dequantize) pairs to avoid numerical overflow on large scale values. Key changes include:

• Introduction of the qdq_scales_fix transformation pass (header, implementation, and protobuf utilities)
• Invocation of the new scale‐propagation pass in backend_manager.cc for GPU when the QDQ optimizer is enabled
• Build updates to link the ONNX protobuf definitions (onnx_proto) into the OpenVINO provider

Reviewed Changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
onnxruntime/core/providers/openvino/qdq_transformations/qdq_scales_fix.h	Declares the new Transform pass interface
onnxruntime/core/providers/openvino/qdq_transformations/qdq_scales_fix.cpp	Implements graph construction, scale propagation, and removal of QDQ pairs
onnxruntime/core/providers/openvino/ov_protobuf_utils.h	Adds helpers to get/set float data in protobuf tensors
onnxruntime/core/providers/openvino/ov_protobuf_utils.cpp	Defines get_float_initializer_data and set_float_initializer_data
onnxruntime/core/providers/openvino/backend_manager.cc	Calls the new scale‐fix pass for GPU under the QDQ optimizer
cmake/onnxruntime_providers_openvino.cmake	Links onnx_proto to the OpenVINO provider target
onnxruntime/core/optimizer/double_qdq_pairs_remover.cc	Fixes missing dimension in newly created initializer

Comments suppressed due to low confidence (4)

onnxruntime/core/providers/openvino/backend_manager.cc:433

• The modified branching drops the so_share_ep_contexts condition for GPU in the first branch and only checks enable_ovep_qdq_optimizer in the new GPU branch. Please verify that GPU behavior when so_share_ep_contexts is true but the optimizer flag is false is still correct.

  if ((session_context_.device_type.find("NPU") != std::string::npos) &&

onnxruntime/core/providers/openvino/qdq_transformations/qdq_scales_fix.h:14

• [nitpick] Add unit tests for the new Transform pass to cover scenarios with varying thresholds, different network topologies, and multiple QDQ patterns to guard against regressions.

Status Transform(const GraphViewer& src_graph,

onnxruntime/core/providers/openvino/qdq_transformations/qdq_scales_fix.cpp:14

• The code uses std::format in ToString(), but is not included. Add #include <format> to ensure compilation succeeds.

#include <algorithm>

onnxruntime/core/providers/openvino/qdq_transformations/qdq_scales_fix.cpp:67

• [nitpick] Remove these commented-out placeholder lines (//** node_input_name = [], //** node_output_name = []) to clean up dead code and improve readability.

      //** node_input_name = []

[sfatimar]

This PR is being merged only because of urgency.. It is not properly code reviewed so may have some issues.