Fix layout transformer for FusedConv

[fs-eire]

Description

Fix layout transformer for FusedConv.

The current layout transformer will transform FusedConv (kMSDomain) into FusedConv (kMSInternalNHWCDomain) if the EP wants channels_last.

However, kMSInternalNHWCDomain uses OpType Conv for both Conv and FusedConv, so FusedConv (kMSInternalNHWCDomain) is invalid (unregistered op).

This PR fixes this and allows layout transformer change FusedConv (kMSDomain) into Conv (kMSInternalNHWCDomain).

Motivation and Context

[tianleiwu]

Has you done some parity test for such FusedConv conversion? My understanding is that FusedConv might have activation, but Conv does not support that.

I noticed that there was NhwcFusedConv:

onnxruntime/onnxruntime/core/optimizer/nhwc_transformer.cc

Lines 102 to 103 in a8fb786

	OpKernelRegistryId nhwc_conv_fp16{
	"NhwcFusedConv", kMSDomain, 1, {{"T", {DataTypeImpl::GetTensorType<MLFloat16>()}}}};

That could impact the layout transformer as well.

[fs-eire]

Has you done some parity test for such FusedConv conversion? My understanding is that FusedConv might have activation, but Conv does not support that.

I noticed that there was NhwcFusedConv:

onnxruntime/onnxruntime/core/optimizer/nhwc_transformer.cc

Lines 102 to 103 in a8fb786

	OpKernelRegistryId nhwc_conv_fp16{
	"NhwcFusedConv", kMSDomain, 1, {{"T", {DataTypeImpl::GetTensorType<MLFloat16>()}}}};

That could impact the layout transformer as well.

Specifically, the Conv in opset kMSInternalNHWCDomain supports activation and this is one of the reason why there is no FusedConv in opset kMSInternalNHWCDomain

[tianleiwu]

FusedConv is not in the list of GetCUDALayoutSensitiveOps for CUDA:

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

Lines 35 to 48 in 1ef3044

	const std::unordered_set<std::string_view>& GetCUDALayoutSensitiveOps() {
	static std::unordered_set<std::string_view> cuda_nhwc_ops = []() {
	return std::unordered_set<std::string_view>{
	"BatchNormalization",
	"Conv",
	"ConvTranspose",
	"GlobalMaxPool",
	"MaxPool",
	"GlobalAveragePool",
	"AveragePool",
	"GridSample",
	"DepthToSpace",
	"SpaceToDepth",
	"LRN"};

So this change shall be fine for CUDA.