Fix marginal linearization in the context of perfect_foresight_with_expectation_errors_solver with homotopy

If a solution corresponding to 100% of the shock can’t be found in the first
informational period, marginal linearization will be performed to extrapolate a
solution.

However, in subsequent informational periods, this extrapolated solution cannot
be used for the initial conditions of endogenous variables, because the initial
conditions are not a true solution of the nonlinear model. For those subsequent
informational periods, the correct approach is to compute the two solutions
needed for marginal linearization using as initial conditions the values
obtained in the same two solutions for the previous informational
periods (stored as oo_.deterministic_simulation.{sim1,sim2}).