cudatoolkit dependency breaks pointops installation with new drivers

[alexrosen45]

Installing cudatoolkit 11 in a conda environment (line 14 in setup shell script)

conda install pytorch=1.9.0 torchvision cudatoolkit=11.1 -c pytorch -c nvidia -y

doesn't work with CUDA 12 drivers. This causes a lot of errors in different issues: #56, #49, #44, etc.

It turns out that the pointops installation works fine with a CUDA 12 toolkit after prepending its system wide installation to PATH

export PATH=/usr/local/cuda/bin:$PATH

or fresh installing the newest (driver compatible) version within the environment

conda remove cudatoolkit
rm ~/anaconda3/envs/pt/bin/nvcc
conda install cudatoolkit # newer version

You can also run nvcc --version to check that it matches the major version in nvidia-smi.

My question is what relies on an older version of the CUDA compiler? If it's a mistake, you should probably remove it.

Edit: I realize it's not really a mistake because you put Ubuntu: 18.04 or higher and CUDA: 11.1 as prereqs, but telling people to use a specific distro and install new nvidia drivers (which is already annoying enough on ubuntu) is a bit crazy. Maybe just put which versions of Ubuntu you tested your project on and require CUDA 11 or higher (or, at least, any CUDA 11 toolkit since it's backward compatible with minor driver versions: https://docs.nvidia.com/deploy/cuda-compatibility/).