FIX: delayed ARMA impulse response and simulation when p > q

[nicoloangileri]

ARMA.impulse_response and ARMA.simulation are shifted by p - q periods when the AR polynomial is longer than the MA polynomial. This includes every pure AR(p) with p >= 2 built with the default theta=0.

>>> import quantecon as qe
>>> qe.ARMA([0.5, -0.2]).impulse_response(5)
array([ 0.  ,  1.  ,  0.5 ,  0.05, -0.075])   # main
array([ 1.   ,  0.5  ,  0.05 , -0.075, -0.0475])   # this PR

The docstring says psi[0] is unity, and statsmodels' ArmaProcess(...).arma2ma gives the second line. qe.ARMA([0.5, 0.1, 0.05]) is delayed by two periods, and simulation starts with p - q exact zeros for the same reason.

The cause is in set_params: (ma_poly, ar_poly, 1) is passed to scipy.signal.dimpulse and dlsim, which read the coefficients as a transfer function in positive powers of z. When ar_poly is longer, the numerator has a lower degree and the system gets len(ar_poly) - len(ma_poly) pure delays. set_params already padded ar_poly when ma_poly was longer; it now also pads ma_poly with zeros, so both polynomials have the same length. spectral_density and autocovariance use freqz, which is not affected by trailing zeros, and their output is unchanged.

The existing tests use theta = np.zeros(3) with three AR coefficients, so the polynomials already had the same length there. The new tests check, for AR(1), AR(2), AR(3), ARMA(2, 1), ARMA(1, 2) and the existing test model:

• impulse_response against the MA(inf) coefficients from the recursion psi_j = theta_j + sum_i phi_i psi_{j-i};
• seeded simulation against sum_j psi_j u_{t-j} built from the same shocks.

Both fail on main and pass here. The full test suite passes locally (745 tests).

Related to #888 (test coverage of the simulation and impulse-response paths in _arma).