Flatten the trained weights into one vector, recording where each block sits
Source:R/stan.R
stan_pack.RdStan has no closures and no struct types, so everything the generated
function needs arrives as a single vector w and is sliced back out inside.
Matrices are flattened column-major, which is exactly what Stan's
to_matrix() expects, so the two sides agree without any transposition.