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

[Fangrui Song via llvm-commits]

https://github.com/MaskRay approved this pull request.

LGTM.
The internal Halide ARM target still passes with this patch.

@davemgreen might want to check the naming of the test llvm/test/CodeGen/ARM/div-by-constant-to-mul-crash.ll
I think it's too specific to use a "*-crash" to name the test.
One of the existing `*div*.ll` tests might be better.

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