Explicitly pinning at CUDA x.x.x, based on the current status of CUDA 10.1?

[leofang]

This is a follow-up from conda-forge/staged-recipes#9959 (comment), in which we realized CuPy does not yet support CUDA 10.1 Update 2 (10.1.243).

I have four questions:

1. When pinning at 10.1, currently which exact version of it is pinned? Note that there are three 10.1's:

• 10.1: 10.1.105
• 10.1 Update 1: 10.1.168
• 10.1 Update 2: 10.1.243

2. The Docker image for 10.1 was found to be 10.1.243, but I noticed the conda-forge channel only provides up to 10.1.168. Wouldn't there be potential conflict?
3. In light of this notorious versioning semantics by Nvidia for 10.1, can we be more explicit for the CUDA versions? I feel one day we would run into some problems if just asking for a major.minor pinning...
4. This is for myself as I'm still learning the packaging process as a new maintainer: Is there a cudatoolkit-feedstock? I can't find one...I would like to know how its recipe is set up.

Thanks!

cc: @jakirkham

[jakirkham]

This will make more sense if we can address this suggestion.