fix(garch): [T074] GARCH-family QMLE Bollerslev–Wooldridge sandwich SEs by default (#173)

GARCH/EGARCH/GJR-GARCH standard errors were computed ONLY from the inverse
observed-information Hessian of the Gaussian negloglik. That likelihood is a
quasi-likelihood; under fat-tailed returns the information-matrix equality fails
and the inverse-Hessian covariance is inconsistent (it understates the true
sampling dispersion), with no way to request robust SEs.

- New shared helpers in src/arch/estimation.jl: `_volatility_loglik_contribs`
  (per-obs Gaussian log-lik ℓ_t, Dual-typed so ForwardDiff flows through the
  variance recursion) and `_qmle_sandwich_cov(H,S) = H⁻¹(S'S)H⁻¹`.
- New per-obs contribution builders `_garch/_egarch/_gjr_loglik_contribs` (repeat
  the negloglik unpack+filter, skip the scalar stationarity penalty — inactive at
  the stationary MLE).
- All three `stderror` methods gain `cov_type::Symbol=:robust`: :robust (aliases
  :qmle/:sandwich/:bw) builds S via ForwardDiff.jacobian of the contributions and
  returns the BW sandwich; :hessian (alias :opg_hessian) keeps robust_inv(H)
  (exact old behavior); bad cov_type → ArgumentError (NaN fallback kept for a
  genuinely singular Hessian). The delta-method scaling loop is unchanged (only
  the source covariance changes), so vcov/confint/report/show inherit :robust.
  No struct field added — stderror recomputes H from m.y, so the kwarg fully
  controls behavior. (Per-call Hessian recompute perf-nit = #189 SUB-5, deferred.)

Tests (TDD, test/volatility/test_volatility.jl): backward-compat (:hessian
reproduces the old inverse-Hessian formula to rtol 1e-8); default==:robust; the
transform-space sandwich is symmetric & PSD; correct-spec Gaussian n=4000
robust/hessian α,β ratios in [0.6,1.6]; fat-tail t(5) divergence (robust α,β SEs
LARGER than hessian, rel >5%); Monte-Carlo dispersion oracle (mean robust SE
closer to std(α̂) than the hessian SE, which understates it); EGARCH/GJR smoke.
All pre-existing volatility tests stay green (StatsAPI compliance 18/18).
volatility.md: cov_type note + API-table row; verify_examples 1/1.

Ref: Bollerslev & Wooldridge (1992). Reliability report finding M-?? (T074).
