Speed up MAST

The fast path is an APPROXIMATION (explicitly not bit-exact; documented in CHANGELOG/final_audit). Four namespace overrides on MAST internals: (1) fast_validateState strips IRLS sanity checks down to the convergence test; (2) fast_bgf no-ops family$validmu/$valideta and loosens the IRLS epsilon from 1e-8 to 1e-5 for ALL families and ALL fits; (3) fast_updateState by default delegates to MAST's original updater (the native C++ cpp_updateState path is gated OFF behind getOption('autozyme.mast.native_update_state') / env AUTOZYME_MAST_NATIVE_UPDATE_STATE=1, because it segfaults under package attest) -- the C++ kernel only handles binomial+logit and gaussian+identity with all(state$good)==TRUE; (4) fast_lrTest_hybrid: a two-pass hybrid that computes a cheap Wald statistic for every gene (chisq df=1 for cont/disc, df=2 for hurdle) then runs the EXACT upstream lrTest only on the top-K=200 genes ranked by Wald hurdle p-value, overwriting those K rows with exact LR results. lrTest fast path activates only for hypothesis of class CoefficientHypothesis or a length-1 character that resolves to exactly one matching column name in coefC. Correct/accepted regime as benchmarked: bayesglm method, a single-coefficient (1 df per component) contrast on a model with discrete+continuous (hurdle) components, where DE conclusions are driven by top-ranked genes (top-100 jaccard ~1, pearson on coef and -log10 p ~1; max coef deviation ~1e-4). The loosened 1e-5 epsilon and stripped validators apply unconditionally to all genes/families on the IRLS path.