Disable no_unique_address for NVCC and CUDA 12.9

[tpadioleau]

Following #434, I need to disable [[no_unique_address]] to make the Kokkos tests pass on a CUDA 12.9 environment on a CEA computer.

With [[no_unique_address]] I have the following errors:

[ RUN      ] cuda_DeathTest.mdspan_space_aware_accessor_invalid_access_from_device
/home/TP026749/kokkos/core/unit_test/TestSpaceAwareAccessorAccessViolation.hpp:31: Failure
Death test: { Kokkos::parallel_for(Kokkos::RangePolicy<ExecutionSpace>(s, 0, 1), *this); Kokkos::fence(); }
    Result: failed to die.
 Error msg:
[  DEATH   ] 
[  FAILED  ] cuda_DeathTest.mdspan_space_aware_accessor_invalid_access_from_device (147 ms)
[ RUN      ] cuda.mdspan_minimal_functional
/home/TP026749/kokkos/core/unit_test/TestMDSpan.hpp:46: Failure
Expected equality of these values:
  errors
    Which is: 2
  0
[  FAILED  ] cuda.mdspan_minimal_functional (0 ms)
[ RUN      ] cuda.mdspan_space_aware_accessor
cudaStreamSynchronize(stream) error( cudaErrorIllegalAddress): an illegal memory access was encountered /home/TP026749/kokkos/core/src/Cuda/Kokkos_Cuda_Instance.cpp:157
Backtrace: ...

[tpadioleau]

@dalg24 may I get your review please ?

[dalg24]

Deferring to Christian who owns that code and who should be able to reproduce without fighting with machine access.

[crtrott]

Hm I don't even get there. For me the compiler crashes.

[ 22%] Building CUDA object tests/CMakeFiles/test_mdspan_at.dir/test_mdspan_at.cpp.o
nvcc warning : Support for offline compilation for architectures prior to '<compute/sm/lto>_75' will be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).
nvcc error   : '"$CICC_PATH/cicc"' died due to signal 11 (Invalid memory reference)
nvcc error   : '"$CICC_PATH/cicc"' core dumped
make[2]: *** [tests/CMakeFiles/test_mdspan_at.dir/build.make:80: tests/CMakeFiles/test_mdspan_at.dir/test_mdspan_at.cpp.o] Error 139
make[1]: *** [CMakeFiles/Makefile2:449: tests/CMakeFiles/test_mdspan_at.dir/all] Error 2

[crtrott]

This change fixes the compiler crash I see with CUDA 12.9 and GCC 11.5 as host compiler. Testing then passes. I don't see issues with CUDA 12.8.

[tpadioleau]

thanks!