Standardizes theta, declares x's center/scale, and accumulates body
(an expression for one observation's log density, in terms of x[n]) over
rows(x) before applying the jacobian once at the end. precompute is the
one place estimators differ: linear_gaussian and the MDN can build their
conditional distribution once, outside the loop, because it depends on
theta alone; MAF has nothing to hoist, so its caller leaves this at the
default. Mirrors de_log_lik_iid() (R/likelihood.R), which sums the same
per-observation log density on the R side; the two have to agree.