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

[Christian Sigg via llvm-commits]

chsigg wrote:

> do you have a pointer to what you have in mind here?

This environment variable controls lazy loading of kernels in the CUDA driver:  
https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#lazy-loading

The documentation is not consistent whether to use `CUDA_MODULE_LOADING` or `CUDA_MODULE_DATA_LOADING` (see e.g. [here](https://forums.developer.nvidia.com/t/the-kernel-warm-up-phase-in-gpu-packet-processing-example-and-concurrent-kernels/326034)). I checked that the former works though: with two different kernels, the test fails without env variable setting and passes with `CUDA_MODULE_LOADING=EAGER`.

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