Clarify that sampling statements vs. target += may not always preserve sampling behavior

Section 7.4 of the Stan reference manual says that using the sampling statement y ~ normal(mu, sigma) and incrementing the target probability directly with target += normal_lpdf(y | mu,sigma) will “both lead to the same sampling behavior” but the sampling statement drops constant terms from the log probability function when adding to the target.

But I believe that it is not necessarily true that you get the same sampling behavior if the sampling statement's execution is conditional on parameter values. E.g. consider the snippet

if(b) {
  y ~ normal(mu, sigma)
}
else {
  target += normal_lpdf(y | mu,sigma)
}

if b depends upon parameter values, then this does not give the same sampling behavior as

target += normal_lpdf(y | mu,sigma)

because in the first snippet we drop the extra additive constant terms only on the subsets of parameter values which cause b to evaluate to true.

Of course, if b depends upon parameter values, there are likely other practical issues with inference for such a model. Nevertheless, it might be good to add a caveat or re-word this statement in the documentation.

(See also discussion here: https://discourse.mc-stan.org/t/target-vs-sampling-does-dropping-constants-preserve-sampling-behavior/29118/)

Start with Section 7.4 of the Stan reference manual and compare the two code examples described in the issue. Read the linked Discourse discussion to understand the caveat about parameter-dependent conditionals. Done means the documentation accurately explains when sampling statements and target increments do not preserve sampling behavior.