Update CUDA Image without tensorflow, make tensorflow CUDA image based on this non-tensorflow CUDA image

[olevski]

See #126 (comment)

[gavin-k-lee]

@olevski can you summarise what the differences/pros/cons are in having a cuda image without tensorflow are? Cuda is just software that allows you to use GPUs, but tensorflow is afaik a software framework right? Alternatives to tensorflow are things like pytorch?

I'm asking just in the context of seeing what the best way to build a custom renku-compatible, gpu-compatible container for DeepLabCut.