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

Benchmarks: refactored logging to write directly into a file instead of an intermediate stringstream

Fixes #62

Moved common headers from Benchmarks to TNL/Benchmarks so they can be used in other projects more easily

The performance on 16 cores is about the same as before, but it takes
significantly less tasks so it should be much faster on less cores.

The build time improved from 4:48 to 1:43 on 16 cores, so the efficiency
is not great, but it's still better than nothing. Maybe the benchmarks
could be reorganized into more independent tasks that could be
explicitly instantiated more efficiently (even if there would be less
explicit instantiations eventually).

Instead of writing the whole log as a valid JSON object, it is much
easier to write one JSON object per line. This can be processed in
Python with the Pandas package using the following code:

df = pandas.read_json(open("foo.log", "r"), orient="records", lines=True)

Also removed the outputMode parameter from the Benchmark and Logging
classes, since it was pretty useless.

Benchmarks: performed loops are returned via the BenchmarkResults struct, base-time via the getBaseTime method