Build anvl-cpu as multi-arch (amd64 + arm64)

[sebffischer]

Summary

• Build anvl-cpu for both linux/amd64 and linux/arm64, producing a real multi-arch manifest list at :latest and :release.
• Use native GitHub-hosted runners (ubuntu-latest for amd64, ubuntu-24.04-arm for arm64) — no QEMU.
• Each platform leg pushes by digest to both Docker Hub and GHCR; a separate merge job assembles the manifest list under each tag.
• CUDA workflow is unchanged; the original docker-build composite action keeps its single-arch behavior so cuda.yaml is unaffected.
• Drop the hardcoded x86_64-pc-linux-gnu libPath in .Rprofile; use R.version$platform so it's correct on arm64.

Why

The current image is published as a manifest list with only an amd64 entry. On Apple Silicon, docker pull ghcr.io/r-xla/anvl-cpu:latest fails with no matching manifest for linux/arm64/v8, which breaks devcontainer up since the CLI's pre-pull step does not honor --platform=linux/amd64 from build.options. Publishing a real arm64 entry fixes that and lets the dev container run natively on M-series Macs.

Notes for reviewers

• arm64 R packages will fall back to source compilation (Posit PPM doesn't ship arm64 Linux binaries for Noble). First daily build will be slower; subsequent builds use per-platform buildcache-linux-arm64 cache.
• New composite actions:
  • .github/actions/docker-build-by-digest/
  • .github/actions/docker-merge/

Test plan

• Open the PR and let the workflow run on pull_request (push disabled). Both build-latest legs should complete on their respective runners; merge-latest is skipped on PR.
• After merge to main, verify docker buildx imagetools inspect ghcr.io/r-xla/anvl-cpu:latest lists both linux/amd64 and linux/arm64.
• On an Apple Silicon host, docker pull ghcr.io/r-xla/anvl-cpu:latest (no --platform) succeeds.

🤖 Generated with Claude Code