Should get better memory locality by arranging loops from most significant to least

https://github.com/dotnet/machinelearning-samples/blob/fb06c72c512b3d007cabc0c01b3136701cb127d2/samples/csharp/end-to-end-apps/ObjectDetection-Onnx/OnnxObjectDetection/OnnxOutputParser.cs#L140

In ParseOutput(), by simply iterating by box, they y, then x, rather than x, then y, then box, memory accesses within the tensor should be more localized, which should work with caching better.