Some cuDNN modules on the system are just listed as CUDA_MAJOR_VERSION.X. Thus, if no cuDNN module is found for the exact CUDA version, we then look for a .X module.

When searching for a cuDNN module, the output was not redirected to stdout.

module avail prints its output to stderr. In interactive shells, it
seems this stderr is automatically redirected to stdout, so grep
would pick it up. This is not the case in batch jobs, so we manually
redirect the stderr output to stdout.

This hopefully makes it easier to spot what's an actual step in the
and what is extra info.

Sometimes jupyter only prints one URL which makes the instruction in
step 6 wrong. Added a more specific instruction and what to do if the
wrong URL is printed by jupyter.