CI: shrink checkouts via shallow / treeless partial clones

[leofang]

Summary

Closes #2091.

15 workflow checkouts used fetch-depth: 0 (full clone with every blob in history). The audit in #2091 showed only 7 actually need the commit graph (setuptools-scm, git merge-base, git worktree), and none need historical blobs. The remaining 8 either do no local git operations or only need tag resolution. This PR applies the three patterns from that audit.

Patterns applied

• Plain shallow clone — fetch-depth: 1 (6 sites, no local git ops)

  • ci.yml ci-vars
  • build-docs.yml
  • release-upload.yml
  • release.yml check-tag
  • release-cuda-pathfinder.yml upload-assets
  • coverage.yml coverage-windows

• Shallow + tags — fetch-depth: 1 + fetch-tags: true (2 sites, lookup-run-id does git rev-parse <tag>)

  • release.yml determine-run-id
  • release-cuda-pathfinder.yml prepare

• Treeless — fetch-depth: 0 + filter: blob:none (7 sites, commit graph needed but not historical blobs)

  • build-wheel.yml (setuptools-scm)
  • ci.yml detect-changes (git merge-base + git diff --name-only)
  • cleanup-pr-previews.yml (git worktree add gh-pages)
  • coverage.yml coverage-linux (setuptools-scm)
  • coverage.yml build-wheel-windows (setuptools-scm)
  • test-sdist-linux.yml (python -m build → setuptools-scm)
  • test-sdist-windows.yml (same)

Should noticeably reduce checkout time, especially on Windows where git operations are slower.

Test plan

• CI passes on this PR (exercises all three patterns: detect-changes, build-wheel, coverage, test-sdist)
• release dry-run / next release exercises release.yml and release-cuda-pathfinder.yml paths (these aren't triggered by PR CI)

-- 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 e673fe1

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2177/
https://nvidia.github.io/cuda-python/pr-preview/pr-2177/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2177/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2177/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[leofang]

CI is fully green, and the checkout-step speedup is significant. Sampling representative jobs from this PR's run (27097717939) against the latest main baseline (27079959023, commit 7ea8a46) — same workflow infra, same runner types:

Job	Runner	Before	After	Savings
ci-vars	ubuntu-latest	133s	1s	−132s (−99%)
detect-changes	ubuntu-latest	108s	4s	−104s (−96%)
Build linux-64 / py3.12	linux-amd64-cpu8	86s	3s	−83s (−97%)
Build win-64 / py3.12	windows-2022	149s	7s	−142s (−95%)
Test sdist win-64	windows-2022	133s	8s	−125s (−94%)

Every checkout is now seconds rather than minutes. Cumulative compute time saved per CI run across the full matrix (6 linux-64 + 6 linux-aarch64 + 6 win-64 builds, 2 sdist jobs, ci-vars, detect-changes, etc.) is on the order of 30+ minutes. Wall-clock on the critical path (ci-vars → detect-changes → build → test) shaves roughly 3 minutes off the start. Windows benefits the most, as expected.

No correctness issues observed: setuptools-scm (git describe), detect-changes (git merge-base + git diff --name-only), and the release-archive flows (git archive, git worktree) all work correctly under the new clone shapes.

-- Leo's bot