Suggested changes for CUDA build

[Xosrov]

These changes were made mostly because the CUDA Dockerfile in the project, while successfully building FFmpeg, didn't recognize nvenc for me. I made some changes that made it work.

Main Changes:

• Use NVIDIA CUDA+Ubuntu as the building stage image instead of Ubuntu base.
• Build and use deviceQuery from CUDA Samples in cuda-ubuntu.dockerfile so compute capability can be detected automatically. Default values are still hard-coded inside the build script, and can be overridden by setting the CUDA_COMPUTE_CAPABILITY env variable.
• Make use of CURRENT_PACKAGE_VERSION more so as to reduce redundancy.
• Add -O2 to nvccflags. Since NVIDIA's own examples do this, I don't think anything is going to break.
• Added more documentation to the README regarding these changes, and updated the example build output to match the the CUDA build and changes in general.

Other Changes:

• Added libzmq support
• Change download link for libsdl (It's not working for me for some reason) to use SDL's Github release page instead
• Change download link for zlib, for the same reason, to use Madler's Github release page instead

The final image containing the binary is still around 650MB.

[markus-perl]

Hi Xosrov, Thank you for the pull request.

Currently the integration tests fail with the following error message:
0.301 CUDA compute capability could not be found... do you have CUDA access in docker?

It should be possible to build the image without an active Nvidia graphics card. Some people need to built the image on a different machine than where the image is later used.

Please update your pull request to pass the integration tests. Then I can merge the changes back to the master branch.

[Xosrov]

Hey Markus, your concerns are valid. I will have to add default values as fallback to prevent build failures.
I will get this done by the end of the week.

[Xosrov]

@markus-perl I've added the default values, but also made some changes to the overall build script. Please run the workflow and let me know what you think.

EDIT: I've updated the original pull request text with more details about the changes

[Xosrov]

@markus-perl The CI problems are fixed.
For OS X, the problem was with the libzmq package using deprecated functions, using a slightly older version fixed the problem.

For the CUDA Dockerfile, the problem was in the naming of the libraries. For example changing libnppc.so -> libnppc.so.12 fixed the issues. I'm not sure why this was having problems in Github CI, even though it worked for me?

Please take a look and let me know what you think. Thanks for the amazing project!

[markus-perl]

Thx for the work. I will merge the changes into the master branch.