Cupy: Support NCCL 2.9.8

[kmaehashi]

Tasks

• [x] Read Release Notes (cuTENSOR / cuDNN / NCCL).
• [x] Update chainer-test to test version. https://github.com/chainer/chainer-test/pull/657
• [x] Update the library installer.
  
  
  
  • [x] Fix cupyx/tools/install_library.py. https://github.com/cupy/cupy/pull/5201
  
  
  • [x] Backport. #5229
  
  
• [x] Documentation
  
  
  
  • [x] Update Requirements section in Installtion Guide https://github.com/cupy/cupy/pull/5201
  
  
  • [x] Backport. #5229
  
  
• [x] ~Implement new features using the version, if any.~

[takagi]

The release note says NCCL 2.9.8 supports CUDA 10.1, but I can not find the link for CUDA 10.1 binary. I keep NCCL 2.8.3 for CUDA 10.1 that looks to be the final NCCL that supports CUDA 10.1. Also, it says NCCL 2.9.8 does not support CUDA 11.1 and 11.2.

[leofang]

Just FYI. On Conda-Forge we build NCCL from source (since it's available on GitHub) for CUDA 9.2 - 11.2. So far no one came to complain, so I assume it still works as of NCCL 2.9.x 😄

[takagi]

The tasks are all done!