CI: resolve branch run by walking HEAD history

[leofang]

Alternative to #2999 for the same underlying issue (#2975), for comparison — not for merging without discussion.

Rewrites branch mode in ci/tools/lookup-run-id to walk the branch's commit history from HEAD and return the first commit whose workflow has a successful run passing the artifact check. The one Actions-API filter used in the hot path is ?head_sha=<SHA>, which is not among the server-side filters community#24626 flags as returning stale slices.

Why this instead of #2999's union-and-cross-check approach:

• gh run list and the server-side branch= / status= / created= filters all drop out — nothing in the hot path hits a filter 24626 names.
• No union, multi-source dedup, or headSha == BRANCH_HEAD preference encoding — walk order is the preference.
• No wall-clock freshness cutoff — a legitimately quiet branch whose last green run is 4 days old is correct, not stale, because that run's SHA is in the branch's commit history.
• The silent-stale mode that caused the Oct 2 nightly regression is impossible by construction: the top of the walk is HEAD, and the "winner" is the newest green reachable from HEAD.

Call cost in the common case:

• Active branch, HEAD has a green run: ~5 API calls (workflow-id, branch-head, 1 commits page, 1 runs lookup, 0 or 1 artifact listing). Cross-check branch CI run lookup with workflow-runs REST API #2999 currently does ~27.
• Quiet branch, N commits back to a green run: ~5+N.
• Scan bounded by LOOKUP_RUN_ID_COMMIT_SCAN_LIMIT (default 200), errors loudly on exhaustion.

Tag mode is untouched (it uses --commit, not an affected filter, and is low call-volume).

Also removes ci/tools/tests/test_lookup_run_id.py — per discussion on #2975, CI tooling that runs on every PR is exercised by CI itself; the orphaned test suite is removed rather than wired in.

-- Leo's bot

[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.

[leofang]

/ok to test b8b2ffe

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-3013/
https://nvidia.github.io/cuda-python/pr-preview/pr-3013/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-3013/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-3013/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[jpascucci-nv]

My ai says "My take: PR 3013’s branch-walk design is the better mitigation. I’d prefer adopting that strategy, but keep the tests (JRP I agree we can lose the test) and probably add a local head_branch == REF check in the head_sha results to avoid accidentally accepting a same-SHA run from another branch. (JRP: I guess this could happen if we merge a branch?)"