[llvm] [NVPTX] Lower allocas to the local address space (PR #204346)

[Tim Besard via llvm-commits]

maleadt wrote:

> I'm a little worried this will have a performance S-curve but I'm not sure what we can do about that and this direction makes sense to me.

FWIW, I didn't notice any additional regressions on our CUDA.jl test suite; it only fixed the one we had seen by upgrading to LLVM 22.

https://github.com/llvm/llvm-project/pull/204346