[CI] Share MLIR cache across GPU jobs

[Phil-amd]

Summary

• use a stable MLIR cache key derived from platform, an explicit environment version, the LLVM revision, and the LLVM build script
• add a prepare-mlir job that builds LLVM once on a cold miss and validates the result with a FlyDSL build
• persist the shared cache only from main; use a one-day workflow artifact to fan out cold PR builds
• make single- and multi-GPU jobs restore the shared cache or consume the prepared artifact instead of compiling LLVM independently

Why

The previous key included the full workflow and the runner label. Unrelated workflow edits invalidated the cache, while each GPU runner and PR produced a separate approximately 2.1 GB entry. Those scoped copies consumed cache capacity and could evict the reusable main-branch cache.

Test plan

• Run actionlint on .github/workflows/flydsl.yaml
• Confirm a cold PR run compiles LLVM only in prepare-mlir
• Confirm GPU matrix jobs consume the prepared artifact and do not compile LLVM
• After merge, confirm main saves the shared cache and a subsequent PR hits it