[llvm] [MISched] Implement `UnanalyzableFrontier`-based DAG construction algorithm (PR #227375)

[Nathan Corbyn via llvm-commits]

cofibrant wrote:

> Could you also provide data on how this changes the generated code?

>From what I've looked at (e.g., a symbol by symbol diff of the `mafft` code), the diff is very small and the vast majority is attributable to scheduling / RA churn. The only potentially performance-sensitive diffs are broken `stp` pairing opportunities. I'm looking into how to mitigate these with a another change to the clustering mutations.

https://github.com/llvm/llvm-project/pull/227375