binary sizes are ballooning, need to prune some architectures. The aim is to get the binaries to be less than 500MB (incl. all deps) to be shippable via PyPI.

• add warning when detected these GPUs at this code block: https://github.com/pytorch/pytorch/blob/4970e733045b1e802c6dfb42c675f29f846be9f4/torch/cuda/__init__.py
• rebuild binaries