Doing this in a destructor is the only place where we can guarantee that
no open file descriptors will be left behind.

I was running a load method on some object in a loop, which was
returning false most of the time, and eventually the process run out of
file descriptors because the file was not closed properly.

The original algorithm used space as one of the separators between types
in template parameter lists, so as a result "A< short int >" was parsed
into a list of ("A", "short", "int").

Implemeting heat equation benchmark.

Moved tnlObject_impl.h back to tnlObject.cpp since function definitions
in header files cause duplicates when building libtnl. The constructor
of tnlObject is __cuda_callable__ so it's empty definition was moved
directly to tnlObject.h to make it visible to nvcc.

Problem with marking tnlObject's virtual destructor as __cuda_callable__
is yet to be solved...

Adding explicit template instantiation.
Moving some _impl.h files into implementation dir.

Finishing new template interface for GMRES.

Adding unit tests for tnlString, tnlObject and tnlArrayManager.

There is gap of missing commits.

Adding edge-skip to grid comparison function.