Add ShouldConvertDataLayoutForOp() API to allow EPs to customize layout sensitive ops

[edgchen1]

Description

Add IExecutionProvider::ShouldConvertDataLayoutForOp() to allow EPs to customize layout sensitive ops. Move existing hardcoded EP-specific logic out of layout transformer code.

Add OrtEp::ShouldConvertDataLayoutForOp to ABI EP API to allow similar customization by plugin EPs.

Motivation and Context

Enable layout sensitive op customization through internal EP interface and the ABI EP API.

[fs-eire]

What is the difference between the newly added interface and the existing GetPreferredLayout()?

[edgchen1]

What is the difference between the newly added interface and the existing GetPreferredLayout()?

GetPreferredLayout returns a single preferred layout value for the whole EP. some EPs have further customization of what ops should be considered layout sensitive. this is hardcoded here currently:

onnxruntime/onnxruntime/core/optimizer/layout_transformation/layout_transformation.cc

Lines 72 to 117 in 3a47bd2

	// handle special cases
	#if defined(USE_JSEP)
	// TODO(fs-eire): Remove special case handing of JSEP once NHWC Resize implementation is fixed
	if (node.GetExecutionProviderType() == kJsExecutionProvider) {
	if (node.OpType() == "Resize") {
	// leave Resize as-is pending bugfix for NHWC implementation. this means the node will remain in the ONNX domain
	// with the original input layout.
	return false;
	}
	}
	#endif
	// NHWC for Resize operator is not implemented on kWebGpuExecutionProvider
	#if defined(USE_WEBGPU)
	if (node.GetExecutionProviderType() == kWebGpuExecutionProvider) {
	if (node.OpType() == "Resize") {
	return false;
	}
	}
	#endif
	// TODO: We don't need to check USE_CUDA || USE_CUDA_PROVIDER_INTERFACE in this function because we're already
	// checking if the node is assigned to the desired EP (e.g., CUDA EP). We should only need to check
	// ENABLE_CUDA_NHWC_OPS.
	#if (defined(USE_CUDA) || defined(USE_CUDA_PROVIDER_INTERFACE)) && ENABLE_CUDA_NHWC_OPS
	if (node.GetExecutionProviderType() == kCudaExecutionProvider) {
	if (layout_sensitive_ops.count(node.OpType())) {
	const auto& cuda_nhwc_ops = GetCUDALayoutSensitiveOps();
	if (!cuda_nhwc_ops.count(node.OpType())) {
	return false;
	}
	}
	}
	#endif
	// TODO: We don't really need EP pre-processor macros in this function because we're already checking if the
	// node is assigned to the desired EP (e.g., QNN EP). There's nothing about this code that absolutely requires
	// conditional compilation.
	#if defined(USE_QNN) || defined(USE_QNN_PROVIDER_INTERFACE)
	if (node.GetExecutionProviderType() == kQnnExecutionProvider) {
	if (node.OpType() == "Upsample") {
	// Upsample is translated to QNN's Resize, which requires the NHWC layout for processing.
	return true;
	}
	}
	#endif

this new interface provides a way for EPs to specify this originally hardcoded logic.

[github-actions]

You can commit the suggested changes from lintrunner.

[skottmckay]