CI: misc cleanup/improvement

[matthewdouglas]

This PR makes various improvements to our CI workflows.

Overall Linux CUDA build jobs seem to be 15-25% faster, under 4 minutes on average. About half the time is spent downloading the Docker container and installing packages.

CPU builds all complete in under 1 minute now as well. Linux/macOS is typically under 15 seconds, and Windows is 35-55 seconds.

The build wheels step is about 35% faster now, completing typically in 40s-1min.

Specific changes:

• Linux CUDA/ROCm builds in Docker are now parallelized with cmake.
• Linux CUDA build skips build dependencies we don't need (gcc-toolset-11 -> gcc-toolset-11-toolchain)
• Linux ROCm build skips unneeded apt-get update
• Windows/macOS native runner builds drop unneeded CMake/Ninja install from PyPI
  • CMake/Ninja versions we need have been available on runners by default for a decent time now.
• Skip sdist in build-wheels step
• Simplify workflow files, remove unused or redundant variables/flags
• Skip trying to further compress wheel artifacts in upload

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.