deps-cuda: make sure cudnn gets installed, pin torch version for CUDA 11

(sry, must commit to master directly – only Docker deploy is affected; current version is broken for all CUDA images)

Makefile

@@ -97,7 +97,7 @@ deps-cuda:
 # works, too:
  shopt -s nullglob; \
  $(PIP) install nvidia-pyindex \
- && $(PIP) install nvidia-cudnn-cu11==8.7.* \
+ && $(PIP) install nvidia-cudnn-cu11~=8.7 \
  nvidia-cublas-cu11~=11.11 \
  nvidia-cusparse-cu11~=11.7 \
  nvidia-cusolver-cu11~=11.4 \
@@ -158,7 +158,7 @@ deps-tf2:
  fi
 
 deps-torch:
- $(PIP) install -i https://download.pytorch.org/whl/cu118 torch torchvision
+ $(PIP) install -i https://download.pytorch.org/whl/cu118 torchvision==0.16.2+cu118 torch==2.1.2+cu118
 
 # Dependencies for deployment in an ubuntu/debian linux
 deps-ubuntu: