LLVM instruction combine error triggered by while loop in reduce kernels

[PhilipFackler]

Parallel reduce is creating an issue with recent LLVM/Julia code.
It throws a LLVM instruction combine error during run.

This error message indicates that LLVM’s optimizer is running into a bug when trying to combine instructions in our parallel_reduce kernel code. This is usually not an error in our logic but rather an artifact of how LLVM is optimizing our parallel_reduce code. The pattern of a while loop with dynamic shared memory can trigger this bug. This issue can be resolved by replacing the while loop with a for loop that computes the reduction in fixed steps.