AMDGPU backend failure for a GEMM with a TT layout

Please see the reproducer [here ](https://godbolt.org/z/MWWfs7b6W)
This works with -O0 flag but starts failing at -O1 flag

Start by running the linked Godbolt reproducer with the AMDGPU backend at -O0 and -O1, confirming that the GEMM with a TT layout fails only at -O1. Trace the differing compiler path through AMDGPU lowering to identify the failing stage; done means the reproducer succeeds at -O1, with regression coverage for this case.

Written by the indexing model from the issue text.