[CI] Make MLIR cache fallback reliable

[Phil-amd]

Summary

• restore the shared MLIR cache only in prepare-mlir
• always upload the prepared tarball as a short-lived run artifact
• make single- and multi-GPU jobs consume that artifact instead of independently restoring the 2.1 GB cache
• preserve the existing behavior of rebuilding once on a restore failure and saving rebuilt caches only from main

Root cause

The lookup-only probe could succeed while a later GPU-job restore failed. In run 30076185312, the signed cache download URL expired at 48.1%; in run 30076858081, the Navi runner reported a cache miss. Because the probe had succeeded, prepare-mlir had skipped artifact upload, so both jobs then failed with Artifact not found.

Test plan

• Run actionlint on .github/workflows/flydsl.yaml
• Confirm prepare-mlir always uploads mlir-install-${{ github.run_id }}
• Confirm all GPU jobs download the prepared artifact without running actions/cache/restore
• Confirm a cache restore failure falls back to one LLVM build in prepare-mlir

Made with Cursor