A bug introduced with the recent max_dt facility prevented the rosenbrock
controller to increase step size in most cases (if max_dt=0). This is fixed
now, and a regression test case has been added.
fixed bug in less_eq_with_sign. equality was not correctly checked for, which
resulted in wrong behavior when the numeric type had
std::numeric_limits<T>::epsilon() == 0.
when state_type == time_type (e.g. 1d odes with state_type = double), some
do_step overloads are disabled due to ambiguities of parameter structure.
However, the initialization of the Adams-Bashforth needs some of those
disabled functions in its initialization. As a fix, I added do_step_dxdt to
the stepper base classes to provide direct access to the required functions
that will not be disabled in the case of state_type == time_type.
