Refactor write_iteration* functions

write_iteration_csv(...) currently requires specifying the log probability lp. This is unneeded for variational inference, and as a hack currently writes 0.0 there: https://github.com/stan-dev/stan/blob/develop/src/stan/variational/advi.hpp#L315-L317

From stan-dev mailing list:

Let's look into refactoring that into less of a hack.  We shouldn't be requiring
lp__ in anything downstream.   Feel free to bring these things up during
design, and we can look into redoing them as we go so we don't wind up with
a hack release that we want to fix later by getting rid of lp__.

- Bob

Example of printing output.csv with cmdstan's print script:

# stan_version_major = 2
# stan_version_minor = 6
# stan_version_patch = 3
# model = bernoulli_model
# method = experimental
#   experimental
#     variational
#       algorithm = meanfield (Default)
# meanfield
#       iter = 10000 (Default)
#       grad_samples = 1 (Default)
#       elbo_samples = 100 (Default)
#       eta_adagrad = 0.10000000000000001 (Default)
#       tol_rel_obj = 0.01 (Default)
#       eval_elbo = 100 (Default)
#       output_samples = 1000 (Default)
# id = 0 (Default)
# data
#   file = bernoulli.data.R
# init = 2 (Default)
# random
#   seed = 3959575278
# output
#   file = output.csv (Default)
#   diagnostic_file =  (Default)
#   refresh = 100 (Default)
lp,theta
0,0.221171
0,0.166458
0,0.297987
0,...
0,...
0,...

Start with src/stan/services/io/write_iteration_csv.hpp and the variational call in src/stan/variational/advi.hpp. Trace the write_iteration* callers and determine how the output path should work without requiring lp. Done means variational output no longer uses a placeholder lp value and the affected callers remain consistent.