Nightly test failure, Cuda builds with Cu* TPLs enabled, Cuda.sparse_sptrsv_double_int_int_TestDevice

[ndellingwood]

Nightly Cuda builds with TPLs enabled (various versions and arch) are failing the sparse_cuda test in the sptrsv check:

cuda/11.2.2 w/ tpls cusparse, cublas, cusolver on Volta70:

07:23:30 [ RUN      ] Cuda.sparse_sptrsv_double_int_int_TestDevice
07:23:30 /home/jenkins/weaver/workspace/KokkosKernels_Weaver_Cuda_Serial_cuda_1122_gcc_831_cusparse_cublas_cusolver_all_scalars/kokkos-kernels/sparse/unit_test/Test_Sparse_sptrsv.hpp:604: Failure
07:23:30 Expected equality of these values:
07:23:30   sum
07:23:30     Which is: -nan
07:23:30   lhs_v[i].extent(0)
07:23:30     Which is: 6
07:23:30 /home/jenkins/weaver/workspace/KokkosKernels_Weaver_Cuda_Serial_cuda_1122_gcc_831_cusparse_cublas_cusolver_all_scalars/kokkos-kernels/sparse/unit_test/Test_Sparse_sptrsv.hpp:604: Failure
07:23:30 Expected equality of these values:
07:23:30   sum
07:23:30     Which is: -nan
07:23:30   lhs_v[i].extent(0)
07:23:30     Which is: 6

Also seen e.g. with cuda/12.2 on A100 (Ampere)

This follows after merge of #2429 , @jgfouca can you investigate?

Reproducer: (weaver rhel8 queue):

bsub -Is -n 1 -q rhel8 -gpu "num=1" bash

source /etc/profile.d/modules.sh
source /projects/ppc64le-pwr9-rhel8/legacy-env.sh
module load cuda/11.2.2/gcc/8.3.1 cmake/3.23.1

$KOKKOSKERNELS_PATH/cm_generate_makefile.bash --with-cuda --with-serial --compiler=$KOKKOS_PATH/bin/nvcc_wrapper --arch=Volta70,Power9 --with-cuda-options=enable_lambda --kokkos-path=$KOKKOS_PATH --kokkoskernels-path=$KOKKOSKERNELS_PATH  --with-scalars='double,complex_double,float,complex_float' --with-ordinals=int --with-offsets=int,size_t --with-layouts=LayoutLeft --with-tpls=cusparse,cublas,cusolver --cxxstandard=17