[llvm] [mlir] [mlir][gpu] Change GPU modules to globals (PR #135478)

[Valentin Clement バレンタイン クレメン via llvm-commits]

clementval wrote:

> I wonder we could move back the logic into `createKernelLaunch()` where it was until now instead of during the translation? We'd still inject the `mgpuModuleLoad` into a global constructor, but we'd do it only if we have `gpu.launch` present, that way a simple translation of a kernel binary does not do that and preserve other runtime/lowering implementations.

That would be nice to keep the translation of gpu.binary to the creation of the global. We can probably work around the current constructor generation in our case. 

https://github.com/llvm/llvm-project/pull/135478