Stan math should have specific type traits only the compiler uses

#1709 is a PR that turns on SoA for math function signatures like vector fma(var, data vector, var). It does this by calling rep_vector on one of the var members of the signature. This creates a size(data vector) SoA matrix in the signature. This is not terrible, but not great either.

I think we should have Stan compiler specific tags and type traits the compiler can use to sneakily tell Stan math things. For instance, we can have a type trait that is just a no-op wrapper around vars to tell stan math to promote the end result to a SoA matrix. So the end C++ we generate would look something like

var_value<Eigen::VectorXd> d = fma(stanc_promote_soa{a_var}, b_vec, stanc_promote_soa{c_var})

stanc_promote_soa would be a minimal class that just holds a reference to a var and tell the stan math type traits that the end result should be a SoA matrix.

This would be a much more efficient impl

Start with PR #1709 and the existing rep_vector use described in the issue, then trace how the compiler generates signatures such as fma(var, data vector, var). Done means defining compiler-specific promotion traits and generating the proposed wrapper-based C++ form, with the resulting Stan Math type behavior verified.