Cleanup spadd (#685, #694)

[brian-kelley]

• Remove "max_nnz_in_result" and its getter/setter from SpADD handle. It was really a duplicate of "c_nnz". Max is a bad name for this since it is not an upper bound, it is always exact. This will need to be propagated to Tpetra_MatrixMatrix_def.hpp when we do promotion.
• Remove team size/vector length stuff from SpADD handle, since SpADD only uses range policy. This was just copied from SpGEMM handle without needing it
• No longer require C's scalar values to be zero-initialized when they are passed in (SpAdd requires output values to be zero-initialized, but this shouldn't be needed #694)
• If A and/or B have unmerged rows (aka repeated column in a row), the resulting C is now always merged. Before, it would produce the correct sum but could only merge one entry from A and one entry from B. Now, it can merge an arbitrary number of entries from A and B (SpAdd doesn't merge entries correctly #685)
• Improve test coverage: test with unmerged inputs (both sorted and unsorted). The test now checks that the result is sorted and merged. Also add a test for a matrix with nonzero dimensions but zero entries.

Note that these are all just enhancements, not bugfixes. I noticed these while trying to track down an EMPIRE issue, where add's MergeEntriesFunctor is crashing after my recent cuSPARSE TPL changes. I think the underlying cause of the bug is somewhere else since cuSPARSE SpMV is completely unrelated to add.

Testing:
#######################################################
PASSED TESTS
#######################################################
clang-8.0-Pthread_Serial-release build_time=336 run_time=173
clang-9.0.0-Pthread-release build_time=136 run_time=77
clang-9.0.0-Serial-release build_time=174 run_time=64
cuda-10.1-Cuda_OpenMP-release build_time=848 run_time=167
cuda-9.2-Cuda_Serial-release build_time=794 run_time=210
gcc-4.8.4-OpenMP-release build_time=117 run_time=64
gcc-7.3.0-OpenMP-release build_time=179 run_time=2532
gcc-7.3.0-Pthread-release build_time=132 run_time=80
gcc-8.3.0-Serial-release build_time=206 run_time=66
gcc-9.1-OpenMP-release build_time=190 run_time=392
gcc-9.1-Serial-release build_time=175 run_time=67
intel-17.0.1-Serial-release build_time=263 run_time=59
intel-18.0.5-OpenMP-release build_time=425 run_time=67
intel-19.0.5-Pthread-release build_time=467 run_time=72
clang-8.0-Cuda_OpenMP-release (test failed) <== sparse_openmp timed out due to machine congestion, but I re-ran it today from the same build and it passed
#######################################################

[lucbv]

This looks good to me, I just have one comment regarding value initialization.

[brian-kelley]

Still good after making the 1-line change:

#######################################################
PASSED TESTS
#######################################################
clang-8.0-Cuda_OpenMP-release build_time=658 run_time=170
clang-8.0-Pthread_Serial-release build_time=210 run_time=141
clang-9.0.0-Pthread-release build_time=148 run_time=71
clang-9.0.0-Serial-release build_time=169 run_time=63
cuda-10.1-Cuda_OpenMP-release build_time=942 run_time=186
cuda-9.2-Cuda_Serial-release build_time=901 run_time=212
gcc-4.8.4-OpenMP-release build_time=122 run_time=67
gcc-7.3.0-OpenMP-release build_time=158 run_time=113
gcc-7.3.0-Pthread-release build_time=185 run_time=76
gcc-8.3.0-Serial-release build_time=146 run_time=63
gcc-9.1-OpenMP-release build_time=215 run_time=151
gcc-9.1-Serial-release build_time=168 run_time=61
intel-17.0.1-Serial-release build_time=255 run_time=58
intel-18.0.5-OpenMP-release build_time=353 run_time=60
intel-19.0.5-Pthread-release build_time=452 run_time=67