[llvm] [NVPTX] Do not run the NVVMReflect pass as part of the normal pipeline (PR #121834)

[Johannes Doerfert via llvm-commits]

jdoerfert wrote:

> I don't think this belongs in the backend, or middle end optimization pipeline. It's really a job for whatever "frontend" is loading the bitcode for final code generation

I get @jhuber6's point about target-specific specialization. There is a benefit if we could do more "library" code IR generation w/o specifying all target details. We kinda do that now, and it broke stuff, but the direction is good. 

What is the downside of multiple specialization runs, with the earlier one(s) not specializing what they do not know for sure?

https://github.com/llvm/llvm-project/pull/121834