CI: Refactor the CI matrix #576

Take a pass at refactoring the CI matrix used for PRs with the goal of adding some grouping to related builds/tests and making the test matrix more configurable to enable CI and nightly test matrix options. Ideally this enables partial matrix expansion on some elements and having a wide variety of coverage while keeping test times under control.

The driving idea for grouping is to have build/test architectures at the top-level, and then group variations on other things such as Python version, local versus wheels CTK, driver versions and GPUs below that.

• Build linux-64
  • py3.9
  • py3.12
  • py3.13
• Build linux-aarch64
  • py3.9
  • py3.12
  • py3.13
• Build win-64
  • py3.9
  • py3.12
• Test linux-64
  • py3.9, 11.8.0, local, l4
  • py3.9, 12.8.0, wheels, l4
  • py3.13, 12.8.0, local, H100
• Test linux-aarach64
  • ...