Rebase TensorFlow GPU image
Currently the TensorFlow GPU image is based on nvidia/cuda
, but now the Jupyter docker stacks also have a TensorFlow image with GPU support, see doc, which would simplify our image a lot. To switch we could either build upon tensorflow-notebook:cuda-tensorflow-* or, if it fails to install from conda-forge after TensorFlow was installed using pip, base upon scipy-notebook and mimic the dockerfile of the cuda-variant to have pip in the end.