[llvm] Reapply "[AMDGPU] Always lower s/udiv64 by constant to MUL" (PR #101942)

[Pierre van Houtryve via llvm-commits]

Pierre-vh wrote:

@MaskRay please check that this is good for Halide now. The test case is based on the reproducer given.
@davemgreen please check the ARM test case

https://github.com/llvm/llvm-project/pull/101942