Move blocking patterns are not correctly handled for `LinModel`

[franckgaga]

Following discussions at darnstrom/LinearMPC.jl#55, it seems that custom move blocking patterns are not correctly handled for LinModel. Here's a MRE:

using ModelPredictiveControl, ControlSystemsBase
using Plots

G = tf( 10, [400,  1])
Ts = 100.0
model = LinModel(G, Ts)
f(x, u, _, p) = p.A * x + p.Bu * u
h(x, _, p)    = p.C * x
nonlinmodel = NonLinModel(f, h, Ts, 1, 1, 1, p=model, solver=nothing)

mpc1 = LinMPC(model; Mwt=[1], Nwt=[0], Hp=30, Hc=[2, 2, 2, 24], transcription=SingleShooting())
mpc2 = LinMPC(model; Mwt=[1], Nwt=[0], Hp=30, Hc=[2, 2, 2, 24], transcription=MultipleShooting())
mpc3 = NonLinMPC(model; Mwt=[1], Nwt=[0], Hp=30, Hc=[2, 2, 2, 24], transcription=SingleShooting())
mpc4 = NonLinMPC(model; Mwt=[1], Nwt=[0], Hp=30, Hc=[2, 2, 2, 24], transcription=MultipleShooting())
mpc5 = NonLinMPC(nonlinmodel; Mwt=[1], Nwt=[0], Hp=30, Hc=[2, 2, 2, 24], transcription=SingleShooting())
mpc6 = NonLinMPC(nonlinmodel; Mwt=[1], Nwt=[0], Hp=30, Hc=[2, 2, 2, 24], transcription=MultipleShooting())

sim!(mpc1, 50, [5], x_0=zeros(mpc1.estim.model.nx), x̂_0=zeros(mpc1.estim.nx̂)) |> plot |> display
sim!(mpc2, 50, [5], x_0=zeros(mpc2.estim.model.nx), x̂_0=zeros(mpc2.estim.nx̂)) |> plot |> display
sim!(mpc3, 50, [5], x_0=zeros(mpc3.estim.model.nx), x̂_0=zeros(mpc3.estim.nx̂)) |> plot |> display
sim!(mpc4, 50, [5], x_0=zeros(mpc4.estim.model.nx), x̂_0=zeros(mpc4.estim.nx̂)) |> plot |> display
sim!(mpc5, 50, [5], x_0=zeros(mpc5.estim.model.nx), x̂_0=zeros(mpc5.estim.nx̂)) |> plot |> display
sim!(mpc6, 50, [5], x_0=zeros(mpc6.estim.model.nx), x̂_0=zeros(mpc6.estim.nx̂)) |> plot |> display

that should all lead to identical results, but it is not the cases for all the controllers based on a LinModel. There is clearly a bug in init_predmat and init_defectmat methods for LinModel. I will add a test that do this comparison once debugged.