Allow truncation syntax in vectorized sampling statements

[WardBrian]

Closes #788. This allows statements like

  y ~ normal(mu1, sigma1) T[L, U];

In place of

  for (n in 1 : N) {
    y[n] ~ normal(mu2, sigma2) T[L, U];
  }

We previously translated the loop into (MIR syntax):

for(n in 1:N) {
    if((y[n] < L)) 
        target += FnNegInf__(); 
    else if((y[n] > U)) 
        target += FnNegInf__(); 
    else 
        target += PMinus__(log_diff_exp(normal_cdf_log(U, mu2, sigma2), normal_cdf_log(L, mu2, sigma2)));
    target += normal_lupdf(y[n], mu2, sigma2);
}

We now turn the vectorized version into

if((min(y) < L)) 
    target += FnNegInf__(); 
else if((max(y) > U)) 
    target += FnNegInf__(); 
else 
    target += PMinus__((log_diff_exp(normal_cdf_log(U, mu1, sigma1), normal_cdf_log(L, mu1, sigma1)) * size(y)));

target += normal_lupdf(y, mu1, sigma1);

NB the calls to min, max, and the multiplication by size(y). This formulation is based on #788 and the linked thread, but also to me seems more or less necessarily equivalent to the for loop.

Submission Checklist

• [x] Run unit tests
• Documentation
  • [ ] If a user-facing facing change was made, the documentation PR is here: TBD

Release notes

Vectorized sampling statements can now be used with the truncation syntax T[ , ]

Copyright and Licensing

By submitting this pull request, the copyright holder is agreeing to license the submitted work under the BSD 3-clause license (https://opensource.org/licenses/BSD-3-Clause)

[bob-carpenter]

I agree this is going to be super useful. And thanks for catching the alternative vectorization cases, @nhuurre.

[WardBrian]

@nhuurre great catches!

My latest two commits:

1. Make it so the truncation is only multiplied by size(y) if none of the later args are vectors
2. Generate a for loop in the T[U, L] vector case.
3. move the lupdf statement above this logic. I believe this is sound (they're both just accumulation statements), my reasoning is that item 2 here could generate a weird error if size(mu) != size(sigma), but the normal_lupdf call will necessarily check this and produce a nice error if it is not true.

[WardBrian]

Thanks @nhuurre. I just circled back to improve the error message which will show if a vectorized lcdf is not available and add a test for that.

I'll merge after the tests pass.