It seems at some point compiling with CUDA got removed from the package, but I can't find reference for why or when. The NVDEC/NVENC things are compiled in fine. Compiling cuda in is needed to enable additional CUDA-based filters that allow video transcoding to be kept in GPU. For example, yadif_cuda. See also https://bugs.gentoo.org/695210 for an additional filter that can be enabled if the system has CUDA
it's masked
We were discussing this on IRC, the reason for the stable mask of "all" USE=cuda is because it tend to pull in nvidia-cuda-* packages which are not stable. However USE=cuda on ffmpeg does not pull this, it just need NVPTX target on llvm. I feel it needs to be exempted from the mask, not that I've investigated further.