Fix using the GPU

[mikegerber]

This is probably the same issue I had while training: Something with the CUDNN path.

In addition I'd like a parameter to

• force GPU usage (fail if not available - less surprises that way)
• force CPU usage

Currently it just uses the CPU in my configuration (Python 3.9, pyenv, virtualenv, CUDA like installed on our servers). At least this answers another question I had: if inference is possible on CPU only.