[llvm] [mlir] [mlir][OpenMP] Add __atomic_store to AtomicInfo (PR #121055)

[Tom Eccles via llvm-commits]

tblah wrote:

> I have made a branch off this PR here: https://github.com/NimishMishra/f18-llvm-project/commits/atomic_allocated_type_size/. Could you check if you are happy with it? I can submit it for review as a separate PR then, since the issue is with all of atomic read/write/update.

Thanks for looking into this. I think we need to perform an actual cast at the fortran level. For example this segfaults when built with omp because the write of `x` overflows `v`:
```
program repo
  complex(8) :: x = 0
  complex v
  integer :: i = 1
  !$omp parallel
    !$omp atomic read
      v = x
  !$omp end parallel
  print *,"ok"
end program
```

Building without openmp there is a `fir.convert` to change `x` into a `complex(4)` before performing the assignment.

https://github.com/llvm/llvm-project/pull/121055