Update ORT to handle explicit OpSchemaRegisterOnce API in ONNX >= 1.18.0 for fluent chaining

[titaiwangms]

ORT was leveraging on implicit api ONNX_NAMESPACE::OpSchemaRegistry::OpSchemaRegisterOnce https://github.com/microsoft/onnxruntime/blob/fe97f8b608bbd9131a8de25b0d6b57a9ee30c388/onnxruntime/core/graph/contrib_ops/contrib_defs.h#L39C10-L39C64 to do fluent chaining style as following

onnxruntime/onnxruntime/core/graph/contrib_ops/collective_defs.cc

Lines 17 to 28 in fe97f8b

	ONNX_CONTRIB_OPERATOR_SCHEMA(AllReduce)
	.SetDomain(kMSDomain)
	.SinceVersion(1)
	.Input(0, "input", "tensors to be reduced", "T", OpSchema::Variadic)
	.Output(0, "output", "reduced tensors", "T", OpSchema::Variadic)
	.TypeConstraint(
	"T",
	{"tensor(float16)", "tensor(float)", "tensor(double)"},
	"Constrain to float, float16 and double tensors.")
	.TypeAndShapeInferenceFunction([](ONNX_NAMESPACE::InferenceContext& ctx) {
	propagateShapeAndTypeFromFirstInput(ctx);
	});

However, the api ONNX_NAMESPACE::OpSchemaRegistry::OpSchemaRegisterOnce will be explicit starting from ONNX==1.18.0.

The PR: onnx/onnx#6378
The change: https://github.com/onnx/onnx/blob/9e379bdd51be054c89b29386f17a2fd731ce190f/onnx/defs/schema.h#L984

It's currently avoid from cmake/patches/onnx.patch within #24449