Build issues when multiple torch versions are present

[alihassanijr]

We read _GLIBCXX_USE_CXX11_ABI in setup.py, but find libtorch and header dirs, and handle linking in cmake, which itself starts a python subprocess to find torch.
We stopped using torch's cmake because it frequently breaks our flow on things that have nothing to do with natten, like Caffe.

We should first resolve this inconsistency, and either move torch header paths, etc to setup.py, or everything into cmake.
The former sounds like the "correct" way of doing it, but we're also leaving cmake in the dark.

We should also try to figure out if we can get Pypi/python to stop installing another version of torch, when users bring their own, without having to tell users to use --no-deps and --no-build-isolation.