fix(gmm): [T073] SMM optimal weighting from per-observation influence contributions (#172)

The SMM moment covariance Ω was built by evaluating the (demeaning) moment
function one row at a time, so every contribution row was identically zero, Ω
collapsed to the 1e-8·I regularization fallback, the two-step "efficient"
weighting degenerated to ~1e8·I, and the reported SEs (D'WD)^{-1} were invalid.
This propagated into estimate_dsge(...; method=:smm).

- New exported `autocovariance_moment_contributions(data; lags)` returning the
  n×q matrix of per-observation contributions whose column-mean is EXACTLY
  `autocovariance_moments(data; lags)` (var-cov block i<=j, then per-lag diagonal
  block; t<=lag entries are 0 to reproduce the /n divisor).
- `_moment_covariance` now consumes a contributions matrix directly;
  `smm_weighting_matrix`/`smm_data_covariance` take a `contributions_fn` (n×q),
  not the mean moments_fn. Ω is the HAC (Bartlett, Newey-West 1994 auto-bw)
  long-run covariance of the contributions.
- W = Ω⁻¹ via safe_cholesky (minimal jitter only if not PD); removed the blanket
  1e-8·I fallback. Bandwidth 0 ⇒ auto-select (now well-defined).
- `estimate_smm` gains `contributions_fn::Union{Nothing,Function}=nothing`;
  two-step needs it for a valid Ω (warns + falls back to identity if absent).
  DSGE-SMM threads a default `d->autocovariance_moment_contributions(d;lags=1)`.
- Also fixed a pre-existing SE bug masked by the degenerate weighting: D is
  differentiated w.r.t. the CONSTRAINED θ, so vcov=(D'WD)^{-1}/n is already in
  θ-space; the extra `J_transform·vcov·J_transform'` double-counted the bounds
  map and deflated SEs by (∂θ/∂φ)². Removed it — SMM SEs now match Monte-Carlo
  dispersion (AR(1) ρ̂: reported 0.040 vs MC 0.042).

Tests (TDD, analytic + statistical oracles): mean-identity to 1e-12; Ω full-rank
& W≠scalar·I & W·Ω≈I; iid-Gaussian variance-moment long-run var → 2σ⁴=2.0;
SMM SE ≈ MC dispersion (rtol 0.35); DSGE-SMM vcov finite/PD.
test_smm.jl 73/73, test_gmm_ext_coverage(GMM) 6/6, test_dsge.jl 1387/1387.

Ref: Ruge-Murcia (2012); Lee & Ingram (1991); Duffie & Singleton (1993);
Newey & West (1994). Reliability report §6.2 M-05.
