[RFC] Use CMake to compile torchvision in setup.py

TorchVision uses setuptools to compile its C++ / CUDA dependencies, which has been working ok for now.

But with the addition of FFmpeg and libpng / libjpeg-turbo as dependencies, and in the interest to simplify the path from using torchvision ops in C++, it might be worth considering unifying the compilation steps in torchvision to aways go through CMake.
This could potentially simplify / make more robust to locate external libraries (like FFmpeg and libjpeg-turbo

This has been first discussed in https://github.com/pytorch/vision/pull/2253#issuecomment-648219377, and has been brought already in the past by @soumith

Non-exhaustive list of what is left to do:

• Optional PNG and JPEG libs (they are required in cmake)
• Verify that FindPNG / FindJPEG work with our setup on CI
• Logic to compile independent libs for image and video
• torchvision_EXPORTS macro definition unconditionally for windows (shouldn't matter too much)
• rocm/hipify logic
• option to build tests for C++ models (can be postponed)

cc @bmanga @andfoy for discussion.

Start at setup.py and the CMake configuration, then review the listed FindPNG/FindJPEG, CI, independent image/video library, Windows, ROCm/hipify, and C++ model test concerns. Done means the compilation path is unified through CMake and the required checklist items are resolved or explicitly deferred.