Replaced TNL::Pointers::SharedPointer with std::shared_ptr in the LinearSolver base class

See merge request !111

This is new in CUDA 11.5, see
https://developer.nvidia.com/blog/reducing-application-build-times-using-cuda-c-compilation-aids/

Code coverage

See merge request !110

Taken from https://github.com/jedbrown/cmake-modules

Updated some old benchmark scripts

See merge request !109

The tnl-matrix-convert and tnl-grid-view tools are not part of TNL
anymore. Both scripts were moved to the tnl-legacy repository.

Otherwise random things may break, since new nvcc versions use C++17 by
default...

- the old v1.10.x does not build with Clang 13
- prevent installing GTest along with TNL, it is used only for tests
  that are not installed themselves

Benchmarks refactoring

Closes #96 and #62

See merge request !107

Benchmarks: removed method newBenchmark and refactored logging of remaining metadata into table columns

The global metadata like hardware information is logged into a separate JSON
file, so the handling of MetadataMap in the Logging classes is not
needed anymore.

Fixes #96