Test performance of numFfmpegThreads with CUDA

[ahmadsharif1]

I don't know how these two interact. From some experiments I ran earlier this year, CUDA performance was best with numFfmpegThreads=1. But we should do these experiments again and that's the case, we should always set 1 thread when using CUDA decoding.

[NicolasHug]

We should also update the docstring of num_ffmpeg_threads to reflect our findings, as originally attempted in #348