The former has been formally deprecated for a long time and in CUDA 10.0
it is finally properly marked as deprecated, so the compiler can print
warnings.

According to benchmarks, there is practically no difference in
performance. Only explicit unrolling is helpful, but that has been
implemented for the general algorithm in Reduction::reduce as well.

Now that error handling is done via exceptions, the return value is free
for more sensible things.

The exception is not used anywhere and does not make much sense.
Functions/methods which are not implemented yet can just throw
std::runtime_error wich a more specific message.

See https://jlk.fjfi.cvut.cz/gitlab/mmg/tnl-legacy

The method was either returning true or throwing an exception, so it can
simply return void or throw an exception.

(Jak má člověk propalovat čas v Ostravě, když mu to školitel rozbil....
git log -L60,60:src/TNL/Meshes/GridDetails/Traverser_Grid3D_impl.h)

- There is an RAII wrapper called ScopedInitializer to handle
  finalization due to C++ exceptions and to avoid calls to Finalize in
  all exit branches.
- Fixed signature of MpiCommunicator::Init to match that of MPI_Init.

Even the methods for specializations without storage have to return
something, which was not the case...

The original was simpler/more readable and we require GCC 5.0 now.

This reverts commit d920a278.

The std::isnan method can be safely used in host code (if the compiler
is not moron). There is no device code which uses isnan so the
TNL::isNan function is useless.

This reverts commit 0eff2f63.