triton and aotriton download their own LLVM builds

[jeroen-mostert]

So triton and aotriton apparently have the same idea as we do, that is reproducible builds, and as part of building they pull a 1 GB blob with LLVM. This download can be very, very slow if you're unlucky (it's the cloud, innit), and moreover it happens twice. It doesn't even appear like this compiler is getting used (?), since all the build commands invoke our own clang. Maybe it's used at runtime, but even then it should probably be using our LLVM...

It would be nice if we could cut out this download, or at least reduce it to once only if it's really unavoidable.

[lamikr]

Adding onnxruntime which does download also many packages. Aat least it would be nice to reuse our version of composable kernel as now I need to patch it with our composable kernel patches.