fix(pathfinder): keep canary probes script-safe

[cpcloud]

Closes #1767.

Summary

• Run the CTK canary probe via python -m cuda.pathfinder._dynamic_libs.canary_probe_subprocess instead of multiprocessing spawn, and add a regression test showing a plain caller script is executed only once.
• Preserve the original isolation requirement behind PR feat(pathfinder): add CTK root canary probe for non-standard-path libs #1595: the canary still runs in a fresh Python interpreter with independent loader state, but it no longer bootstraps through the caller's __main__, so script entrypoints without an if __name__ == \"__main__\" guard do not recurse or fail during multiprocessing startup.
• Anchor the canary subprocess to the parent package's import root so the child resolves the same cuda.pathfinder distribution as the parent. This avoids wheel-based CI importing the checkout tree copy, which can be missing generated files such as cuda.pathfinder._version, and keeps source-vs-wheel behavior consistent.
• Update the header-discovery/release-note wording so the canary path is described as a dedicated subprocess while leaving the shared spawned_process_runner in place for the callable-based isolation paths and real-loading tests that still need it.

Reviewer context

• PR feat(pathfinder): add CTK root canary probe for non-standard-path libs #1595 originally introduced the canary probe as a dedicated python -m ... subprocess.
• During review it was switched to multiprocessing.get_context(\"spawn\") to avoid a forked child inheriting preloaded CUDA libraries from the parent and to keep the probe isolated/predictable.
• This change keeps that same guarantee because the canary still runs in a dedicated module entrypoint in a fresh interpreter; the difference is only that it no longer re-imports the user's script as __main__, which is the failure mode reported in [BUG]: run_in_spawned_child_process fails if used as script #1767.
• While validating the change in wheel CI, the child python -m ... process was found to be resolving cuda.pathfinder from the checkout tree because run-tests pathfinder runs from ./cuda_pathfinder. In that environment the checkout copy can lack generated files such as cuda.pathfinder._version, so the child fails before probing.
• Setting the child cwd to the already-imported parent package's import root makes the subprocess import the same distribution root as the parent, whether the parent came from source or from an installed wheel.

Made with Cursor

[rwgk]

I think this is good to merge with the two tiny suggested changes.

As a follow-on: could it be better to eliminate spawned_process_runner.py, and instead use the exact same approach for creating a fully isolated subprocess also for the tests?

[rwgk]

Awesome, thanks so much!

[github-actions]

Doc Preview CI
Preview removed because the pull request was closed or merged.

[seberg]

That was quick, thanks!