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.

Works fine with g++ 5.5 and nvcc 8.0 in Arch Linux. File a bug report on
Gitlab if you have problems.

Plus the associated revamping of the SolverStarter. Also
ODESolverMonitor has been removed as its functionality has been merged
into the IterativeSolverMonitor.

The call to the nextIteration() method must occur exactly once per
iteration, otherwise incrementing the total iterations counter in
wrong places might cause serious bugs. For example if minIterations is
equal to 1 and the initial residue is smaller than convergenceResidue,
the loop on line 154 in tnlGMRESSolver_impl.h will be entered, but the
m-loop on line 179 will be skipped and on line 263 update() will be
called with empty vectors, which will lead to division by zero.

Implemeting heat equation benchmark.

The condition substituting nextIteration() is different than in the
method itself, because incrementing the iteration counter is avoided.

This will ensure that in verbose mode, the final state of the linear
system solvers is always printed. Previously this was the case only when
refresh rate was equal to 1.

Scripts tnl-compile and tnl-link were added.

Parallel reduction in CUDA fixed.
Boundary conditions in linear system assembler fixed.
Min. iterations for linear solvers fixed.