Adds a cmake option (CODE_COVERAGE) to compile a build with relevant options to GCC and NVCC to generate code coverage profiling.

This can then be used with coverage tools like gcovr or coveralls to generate a detailed coverage summary (e.g. in HTML, JSON, XML, etc.).

Also adds a gcovr.cfg with basic options and a conda dependency on gcovr. Finally, it adds some explicit template instantiations in test files to ensure all template class methods are included in coverage analysis (so coverage % isn't artificially high). (reserved for future PR)

Instructions for generating code coverage

mkdir -p <RMM_ROOT>/build/debug && cd <RMM_ROOT>build/debug
cmake .. -DCMAKE_BUILD_TYPE=Debug -DCODE_COVERAGE=ON
make
ctest ## or ninja
cd <RMM_ROOT>
covr -r . -j 12 -o build/debug/coverage/ build/debug/tests/CMakeFiles/

Note that parallel building should not be used, because --keep is used with .cu files.

With rapids-compose, just set build type to Debug in .env, then

build-rmm-cpp -DCODE_COVERAGE=ON
test-rmm-cpp
covr -r . -j 12 -o build/debug/coverage/ build/debug/tests/CMakeFiles/

Then open build/debug/coverage/coverage_details.html. It looks like this:

Summary of changes required to get good coverage analysis:

• Compile with --coverage,-fprofile-abs-path,-fkeep-inline-functions,-fno-elide-constructors and link with --coverage.
• Compile CUDA code --keep (otherwise gcov gets errors finding NVCC's deleted intermediate-stage files)
• Explicitly instantiate template classes at least once in test source files. This ensures any unused methods in these classes get shown as not covered by tests. Reserved for a future PR to improve code coverage.