[NCCL] - Various improvements

[goliaro]

This PR introduces various improvements in the process of cross-compiling Legion/NCCL:

• Download pre-built NCCL libraries, instead of compiling from scratch, and package the relevant binaries into a tarball.
• Fix issue related to nested folders within compiled Legion / NCCL tarballs
• Precompile one version of Legion for each of the following Python versions: 3.7, 3.8, 3.9, 3.10
• Use earliest patch version of CUDA to prevent conflicts with CUDA version installed on user machine
• Only use major and minor CUDA / Python versions in pre-compiled binary URLs
• Add shellcheck