fix(pvar): [T082] structural Σ for IRF/FEVD from innovation (not level) residuals (#181)

Both PVAR estimators built the structural residual covariance Σ from LEVEL residuals that
retain the unit fixed effect α_i (in y_it = α_i + Σ Φ_j y_{i,t-j} + ε_it the level
residual equals α_i + ε_it). So Σ = Var(α_i) + Σ_true on the diagonal (plus between-unit
cross terms), inflating the innovation scale that drives OIRF/GIRF magnitudes — on a
DGP with Var(α_i)≈1, innovation sd 0.1, the reported OIRF impact was ≈sqrt(1.01) instead
of the true 0.1. (FEVD and all coefficient SEs are scale-invariant / independent of Σ and
were unaffected.)

Fix:
- GMM path: new `_compute_innovation_sigma` recomputes Σ from the SAME transformation used
  for estimation (rebuilding X_full = lags+predet+exog and applying :fd/:fod to Y_eff and
  X_full), then maps the transformed-residual covariance back to the level-innovation
  scale — 0.5 for :fd (E[Δε Δε'] = 2Σ) and 1.0 for :fod (orthonormal Helmert ⇒ covariance
  = Σ exactly). Recomputing from group_Y_levels also makes it correct for System GMM
  (the stacked level moment block is only for identification, never Σ). Also fixes a
  secondary predetermined/exogenous-regressor drop in the old helper.
- FE-OLS path: Σ from the already within-demeaned pool (E_pool = Y_pool − X_pool·Φ'),
  which annihilates α_i — replaces the level-reconstruction loop.

Test: "PVAR Σ removes fixed effects (T082)" — FE-OLS/GMM-FD/GMM-FOD all recover diag(Σ)≈
0.01 (was ≈1.01); max(diag Σ)<0.05 (old code fails); OIRF impact diag ≈ innovation sd 0.1
(not ≈1.0); FD/FOD Σ agree; FEVD rows sum to 1. GMM lags capped in the test to keep
n_inst tractable on the long panel. test_pvar.jl green (Analysis Cholesky identity intact).
Ref: reliability report finding M-15.
