Windows C++ library build issues

🐛 Bug

Thanks for your work in putting up this great software. PyTorch and TorchVision are two amazing tools.

I encountered two different issues when trying to build the C++ library on Windows:

1. The debug build tries to link to the debug python library which most people do not have.
2. The operators registration code was not included in consuming projects.

To Reproduce

Steps to reproduce the behavior:

1. Download the latest LibTorch builds
2. Build and install the vision C++ library with CMake by giving it the proper CMake arguments, e.g. Torch_DIR, etc.
3. Build a small exectuable that lists the registered ops like this:

Credits to @zsef123 for the code below

#include <iostream>

#include <torch/torch.h>
#include "vision.h"

int main(int argc, char* argv[]) {
    auto& ops = torch::jit::getAllOperators();
    std::cout << "torch jit operators\n";
    for (auto& op: ops) {
        auto& name = op->schema().name();
        if (name.find("torchvision") != std::string::npos)
            std::cout << "op : " << op->schema().name() << "\n";
    }
    std::cout << "\n";
    return 0;
}

Expected behavior

A debug build should be possible since it is almost always required on Windows.
Operators should register automatically in consuming binaries.

Environment

• PyTorch / torchvision Version (e.g., 1.7.0 / master):
• OS: Windows 10:
• Installed LibTorch from pytorch.org
• cmake --build .
• Python version: 3.8
• CUDA/cuDNN version: 10.2

Additional context

• The first issue is easily fixed by ignoring the python debug library by adding the /NODEFAULTLIB:python38_d.lib linker flag.
• The second issue is in fact multiple issues. First, if I am not mistaken, since the vision.h file is never included in the vision.cpp the cuda_version definition is considered as a different symbol from the declaration in vision.h. An easy fix, is to add VISION_API in front the definition. However, this is more elegantly fixed by including the vision.h header and changing it so it builds correctly. Second, even with this symbol included in the consuming library, the linker strips the operators registration symbols in a Release build. This is fixed by using the /OPT:NOREF linker flags.

I will submit a PR that fixes all these above issues.

cc @peterjc123 @nbcsm @guyang3532 @maxluk @gunandrose4u @smartcat2010 @mszhanyi

Reproduce the Windows C++ build with CMake and the stated Torch_DIR and LibTorch setup. Inspect vision.cpp, vision.h, and the linker settings for the Python debug library and operator registration. Done means debug builds no longer require python38_d.lib and operators register in consuming Debug and Release binaries.