Add fallback condition to use spmv_native when cuSPARSE won't work

[brian-kelley]

• Improve SpMV unit test:
  • generate random complex values with nonzero imaginary component
  • catch exceptions in spmv
• Add "spmv_native" support to the cusparse spmv file (this calls KokkosKernels spmv)
• Fall back to native for mode 'C' (which cusparse doesn't have), or modes 'T'/'H' for CUDA 9 (fixing cuda.sparse_spmv test failure on Kepler arch with cusparse in Trilinos nightly build #833)
• Fix some minor type casts/checks in spgemm/sptrsv cusparse wrappers - now those builds and tests succeed with all 4 possible combinations of offset/ordinal width

[brian-kelley]

Kokkos-dev2, non-TPL:
#######################################################
PASSED TESTS
#######################################################
clang-8.0-Cuda_OpenMP-release build_time=690 run_time=146
clang-8.0-Pthread_Serial-release build_time=243 run_time=157
clang-9.0.0-Pthread-release build_time=154 run_time=69
clang-9.0.0-Serial-release build_time=147 run_time=58
cuda-10.1-Cuda_OpenMP-release build_time=955 run_time=135
cuda-11.0-Cuda_OpenMP-release build_time=1026 run_time=133
cuda-9.2-Cuda_Serial-release build_time=894 run_time=194
gcc-7.3.0-OpenMP-release build_time=163 run_time=52
gcc-7.3.0-Pthread-release build_time=126 run_time=75
gcc-8.3.0-Serial-release build_time=164 run_time=61
gcc-9.1-OpenMP-release build_time=192 run_time=49
gcc-9.1-Serial-release build_time=174 run_time=53
intel-17.0.1-Serial-release build_time=307 run_time=59
intel-18.0.5-OpenMP-release build_time=426 run_time=52
intel-19.0.5-Pthread-release build_time=431 run_time=75

TPL:
#######################################################
PASSED TESTS
#######################################################
clang-8.0-Cuda_OpenMP-release build_time=705 run_time=146
clang-8.0-Pthread_Serial-release build_time=240 run_time=158
clang-9.0.0-Pthread-release build_time=137 run_time=75
clang-9.0.0-Serial-release build_time=155 run_time=58
cuda-10.1-Cuda_OpenMP-release build_time=1727 run_time=124
cuda-11.0-Cuda_OpenMP-release build_time=1732 run_time=136
gcc-7.3.0-OpenMP-release build_time=170 run_time=51
gcc-7.3.0-Pthread-release build_time=144 run_time=72
gcc-8.3.0-Serial-release build_time=170 run_time=56
gcc-9.1-OpenMP-release build_time=194 run_time=50
gcc-9.1-Serial-release build_time=173 run_time=56
intel-17.0.1-Serial-release build_time=299 run_time=59
intel-18.0.5-OpenMP-release build_time=397 run_time=49
intel-19.0.5-Pthread-release build_time=449 run_time=70

RIDE, non-TPL:
#######################################################
PASSED TESTS
#######################################################
cuda-10.1.105-Cuda_OpenMP-release build_time=713 run_time=149
cuda-9.2.88-Cuda_OpenMP-release build_time=711 run_time=176
cuda-9.2.88-Cuda_Serial-release build_time=692 run_time=231
gcc-6.4.0-OpenMP_Serial-release build_time=238 run_time=177
gcc-7.2.0-OpenMP-release build_time=161 run_time=59
gcc-7.2.0-OpenMP_Serial-release build_time=234 run_time=174
gcc-7.2.0-Serial-release build_time=160 run_time=69
#######################################################
FAILED TESTS
#######################################################
cuda-10.1.105-Cuda_Serial-release (test failed)
#######################################################

TPL:
#######################################################
PASSED TESTS
#######################################################
cuda-10.1.105-Cuda_Serial-release build_time=925 run_time=203
gcc-7.2.0-OpenMP-release build_time=171 run_time=62
gcc-7.2.0-Serial-release build_time=161 run_time=71
gcc-7.4.0-OpenMP-release build_time=176 run_time=59
#######################################################
FAILED TESTS
#######################################################
cuda-9.2.88-Cuda_OpenMP-release (test failed)
#######################################################

Only failures on RIDE were due to #799 (cuda.batched_scalar_teamvector_solve_utv2_double)

[ndellingwood]

Changes look good to me, thanks @brian-kelley !