Use setdefault for internal cache to ensure identity caching

[seberg]

Switch selected cache write sites to setdefault/get-based patterns and keep explicit comments where setdefault is intentionally not used for primitive value caches.
(I just thought having setdefault mentioned might be nice in case of cargo-culting.)

Towards gh-2194; pytest-run-parallel does have a specific test for the GraphMemoryResource change. I could add an identity test for ComputeCapability probably (I doubt it actually matters for it, the use of setdefault just feels like a good pattern to me).

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[seberg]

/ok to test 3497ccf

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2217/
https://nvidia.github.io/cuda-python/pr-preview/pr-2217/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2217/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2217/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[mdboom]

LGTM.

Do we want to add any AGENTS.md content about this pattern being preferred over functools.cache or the non-setdefault caching pattern?

[seberg]

Do we want to add any AGENTS.md content about this pattern being preferred over functools.cache or the non-setdefault caching pattern?

Not sure AGENTS.md has such code styles yet? But if we want to add that I can follow-up.
FWIW, in most cases @functools.cache is nicer, but if it matters to return an idempotent result strictly then it doesn't (at the moment).

However, @rwgk: When scanning for @cache I didn't modify cuda-pathfinder since almost all places were just strings, etc. where it is good.
The one odd-one-out that caught my eye is load_nvidia_dynamic_lib if loading twice from two threads or the function being strictly idempotent can matter, then we need to fix it (I can do that in a follow-up if it seems important to you.)