no tensorflow-gpu on local docker

[suhlrich]

Local docker build can't use GPU for tensorflow. I think it is because of tensorflow/cuda/cudnn version mismatch. We changed tensorflow version likely because we were using a docker image with cuda on aws. When we have to install cuda and cudnn from scratch, we may be able to go back to the versions we use locally.

On my desktop wsl, tensorflow=2.9.3 cuda=11.4 cudnn=8.2

That's not a combination that's been tested: https://www.tensorflow.org/install/source#gpu

@antoinefalisse when you set up the last hpl desktop, maybe try a combination of cuda, cudnn, and tensorflow that works in the above link, or the combination we use locally.

I committed and commented some things in the dockerfile for cudnn. You can uncomment, change the cuda and cudnn versions, and cudnn should be installed when we make

Fortunately tensorflow-cpu seems fine, so this is not urgent.