[llvm] [SeparateConstOffsetFromGEP] Support GEP reordering for different types (PR #90802)

[Benoit Jacob via llvm-commits]

bjacob wrote:

Our downstream project https://github.com/iree-org/iree is getting numerical errors in end-to-end testing on AMD GPU (MI250) / ROCm when integrating llvm-project at this commit. Bisection pointed to this PR and we have verified that reverting this single commit fixes the errors.

Integration with this commit reverted: https://github.com/iree-org/iree/pull/17380

Sample numerical errors with this commit: https://github.com/iree-org/iree/actions/runs/9073149299/job/24930053114#step:11:46 :

```
  expected:
1x4x128x128xf16=[[[1.19141 1.06152 1.05859 1.06543 1.05469 1.04102 1.03223 1.0498 1.03711 1.05469 1.0498 1.05176 1.04004 1.03613 1.03711 1.04102 1.0459 1.05566 1.02832 1.03223 1.02832 1.0166 1.02148 
...
  actual:
1x4x128x128xf16=[[[0.804199 0.852051 0.716797 0.875488 0.828125 0.725586 0.649414 0.885742 0.806641 0.727051 1.00488 0.570801 0.816406 0.820312 0.776367 0.811523 0.744629 0.735352 0.695801 0.701172 
```

-> all elements in these arrays are wrong, but they remain finite values in the same range.

https://github.com/llvm/llvm-project/pull/90802