An autoregressive flow whose per-dimension transform is a monotonic
rational-quadratic spline (Durkan et al., 2019) instead of MAF's affine
shift-and-scale. Splines are far more expressive per layer, which helps on
sharply non-Gaussian posteriors (SLCP, two moons). We reuse the MADE
masking machinery from R/flows.R; each MADE outputs 3K - 1 spline
parameters per dimension (K bin widths, K bin heights, K - 1 interior
derivatives). The spline acts on [-B, B] and is the identity outside
(linear tails), so the standard-normal base distribution is unaffected in
the tails. Note: NSF implementations elsewhere often use coupling layers;
ours is autoregressive — same density family, different conditioning
structure.