[flang-rt] ptxas fatal : Cannot take address of function 'free'

Trying to call _FortranAAssign from the GPU with OpenMP offloading prints an error when flang-rt is built with `-DRUNTIMES_nvptx64-nvidia-cuda_LLVM_ENABLE_RUNTIMES`.

```fortran
program a
integer, dimension(:), allocatable :: xs
integer, dimension(:), allocatable :: ys

!$omp target
xs = ys
!$omp end target
end program
```

```console
$ flang -fopenmp --offload-arch=sm_89 assign.f90
ptxas fatal : Cannot take address of function 'free'
/home/acc/Documents/flang_root/install/bin/clang-nvlink-wrapper: error: 'ptxas' failed
clang: error: nvlink command failed with exit code 1 (use -v to see invocation)
/home/acc/Documents/flang_root/install/bin/clang-linker-wrapper: error: 'clang' failed
flang-23: error: linker command failed with exit code 1 (use -v to see invocation)
```

Looks like there is a workaround for this already in the CUDA/OpenMP build:
https://github.com/llvm/llvm-project/blob/f95ed258260b0349c9bbe8210b0d0726f95a9305/flang-rt/include/flang-rt/runtime/allocator-registry.h#L34-L46

Start with the Fortran reproducer and the flang command using OpenMP offloading to reproduce the ptxas error. Then inspect flang-rt/include/flang-rt/runtime/allocator-registry.h around lines 34-46, where the issue points to an existing CUDA/OpenMP workaround. Done means the reproducer no longer reports that ptxas cannot take the address of free.

Written by the indexing model from the issue text.