fix(dsge): posterior-draw irf/fevd/simulate skip the observation equation

Latent v0.6.2 regression surfaced while regenerating docs plots: the
posterior-draw rebuilds in irf/fevd/simulate(::BayesianDSGE) call
_build_solution_at_theta with ALL endogenous variables as observables and
measurement_error=nothing, so any model with more endogenous variables than
shocks tripped the #139 stochastic-singularity guard (rethrown as non-benign
by design) even though those code paths never use the observation equation.

_build_solution_at_theta gains build_observation::Bool=true; the three
moment rebuild sites pass false and receive (sol, nothing). Estimation,
smoothing, and Bayesian HD paths are unchanged (they genuinely need Z/d/H).

Regression test: 3-endogenous/1-shock model estimated on one observable;
irf/fevd/simulate on the posterior must not throw. test_bayesian_dsge
1010/1010.
