Dynare++: fix computation of IRFs

Since the introduction of the --burn option (in Dynare++ shipped with Dynare
4.3.0), the IRFs reported by Dynare++ were wrong.

The IRFs are computed using a generalized IRF method: the result is
the (average) difference between a simulation with shock and a simulation
without shock. The problem was that the two simulations were not using the same
starting point.

Closes #1634