Clear up parallel processes if processing fails

If our parallel processing fails it looks like we don't have anything to clear up the parallel processes if they fail for some reason leading to error messages of the form:

  Error in `ensure_rstan()`: In order to use `method_bayes()` the `rstan` package must be installed. This can be installed from CRAN by running:
  
        install.packages('rstan')
  
  Please note that for `rstan` to work you need to ensure you have a valid C++ toolchain; for details please see:
  https://github.com/stan-dev/rstan/wiki/RStan-Getting-Started#configuring-c-toolchain
  
  
  Backtrace:
      ▆
   1. ├─base::suppressWarnings(...) at test-reproducibility.R:122:5
   2. │ └─base::withCallingHandlers(...)
   3. ├─rbmi::draws(dat, dat_ice, vars, meth, quiet = TRUE) at test-reproducibility.R:123:9
   4. └─rbmi:::draws.bayes(dat, dat_ice, vars, meth, quiet = TRUE)
   5.   └─rbmi:::fit_mcmc(...)
   6.     └─rbmi:::get_stan_model()
   7.       └─rbmi:::ensure_rstan()
  
  [ FAIL 6 | WARN 0 | SKIP 15 | PASS 1253 ]
  Error: Test failures
  In addition: Warning messages:
  1: In for (call in calls) { :
    closing unused connection 8 (<-localhost.localdomain:11879)
  2: In for (call in calls) { :
    closing unused connection 7 (<-localhost.localdomain:11879)
  3: In for (call in calls) { :
    closing unused connection 6 (<-localhost.localdomain:11879)
  4: In for (call in calls) { :
    closing unused connection 5 (<-localhost.localdomain:11879)
  Execution halted

Start with test-reproducibility.R around lines 122-123 and trace rbmi:::draws.bayes through fit_mcmc, get_stan_model, and ensure_rstan while reproducing the failed parallel run. Done means failed processing leaves no parallel processes or unused connections, and the reported test no longer emits the cleanup warnings.