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

[Ivan Butygin via llvm-commits]

Hardcode84 wrote:

Some drive by comment: It would be nice to have `gpu.launch_func` version which takes func as ssa value + dedicated `gpu.load_module`/`gpu.get_func`, in addition to the existing `gpu.launch_func` which takes symbol ref. This way we can progressive lower thing completely within MLIR, while keeping runtime calls lowering straightforward and delegating kernel lazy loading/caching decisions to the user.

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