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

[Mehdi Amini via llvm-commits]

joker-eph wrote:

> Once that's complete, we remove the gpu runtime and that pass. At the current state, there would be close to zero code duplication.

I'm not sure we want to remove the runtime and the custom translation? I would think this is useful for users that can't (or don't want to) use the LLVM offloading infra for their runtime. 
Or are we confident that the offloading setup is good for all the use-cases that vendors would want?

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