fix(pvar): [T081] one-step FD-GMM includes the Arellano-Bond H band matrix (xtabond2 parity) (#180)

PVAR one-step difference-GMM used an identity-structured initial weight (Σ_g Z_g'Z_g)⁻¹
instead of the Arellano-Bond H band matrix that accounts for the MA(1) error induced by
first differencing. The estimator was consistent but inefficient and did not match
xtabond2 one-step.

Fix: build the efficient one-step weight W₁ = (1/N Σ_g Z_g' H_g Z_g)⁻¹ where H_g is the
r×r band matrix with 2 on the diagonal and -1 on the first off-diagonals (new
`_ab_h_matrix(T, r)`). Applied ONLY for `transformation=:fd`; forward orthogonal
deviations produce homoskedastic errors (H = I), for which Z'Z is already optimal, so the
FOD path is unchanged. W₁ is equation-invariant, so it is now built once (hoisted out of
the per-equation loop) rather than recomputed per equation. `robust_inv(Hermitian(...))`
tolerates near-singular blocks. Two-step re-weights from residuals and still works.

Test: "PVAR one-step GMM Arellano-Bond H matrix (T081)" — (1) H-band structure pin;
(2) EXACT-EQUIVALENCE oracle: the one-step coef reconstructed independently with the H
weight (reusing the package's own _panel_lag_levels / _panel_first_difference /
_build_instruments_fd / linear_gmm_solve helpers) matches estimate_pvar to atol 1e-7;
(3) BUG-EXPOSURE: the H-weighted estimate genuinely differs from the old Z'Z-weighted one;
(4) FOD one-step + two-step still finite. test_pvar.jl green. (Stata xtabond2 is the
reference oracle of record but cannot run in-repo; the H-band form is the textbook AB
weight and the reconstruction pins it exactly.) Ref: Arellano & Bond (1991).
