[llvm] [AMDGPU] Run LowerLDS at the end of the fullLTO pipeline (PR #75333)

[Jon Chesterfield via llvm-commits]

JonChesterfield wrote:

Moving the pass is probably fine - it should be kept adjacent to promotealloca but otherwise doesn't matter much.

The flag is ugly but it'll take some effort to make the pass properly composable, in the sense that it can be run repeatedly or on separate TUs before on the whole module.

Similarly hacky to the flag, we could look for the magic lookup table it builds (has a known name) and skip the pass if it already exists.

https://github.com/llvm/llvm-project/pull/75333