Re-think when the time step is updated

[bangerth]

#856 points out, in a comment, that postprocessors need to use this->get_old_timestep() because the time step is updated between solving and calling postprocessors. This is, at least, confusing.

Indeed, the code looks like this in core.cc:

    // start the principal loop over time steps:
    do
      {
        start_timestep ();

        // then do the core work: assemble systems and solve
        solve_timestep ();

        pcout << std::endl;

        // update the time step size
        // for now the bool (convection/conduction dominated)
        // returned by compute_time_step is unused, will be
        // added to statistics later
        old_time_step = time_step;
        time_step = std::min (compute_time_step().first,
                              parameters.maximum_time_step);
        time_step = termination_manager.check_for_last_time_step(time_step);

        if (parameters.convert_to_years == true)
          statistics.add_value("Time step size (years)", time_step / year_in_seconds);
        else
          statistics.add_value("Time step size (seconds)", time_step);

....

       postprocessor();

This also means that we output the next time step into the statistics file, not the one that was used for the current computations. I think we should re-think the current structure.