Currently, the C++ backend uses "#pragma omp parallel" to emit parallel for loops. We should make this an option under a Target flag and also put in code to call halide_do_par_for with appropriate closures.