Merged and pushed.

I have run a little microbenchmark and the results seem to be slightly in favor of range-based loops, though there is 
some small variation across different tasks:

experiments/issue510/iterator-microbenchmark$ make -j2 && ./benchmark < .../scanalyzer-sat11-strips/p20.pddl/output
...
Running nothing 10000 times: 2.4e-05 seconds

Running standard loop over operators 10000 times: 5.52003 seconds
Running range-based loop over operators 10000 times: 5.51173 seconds

I have uploaded my implementation to 
https://bitbucket.org/jendrikseipp/downward/branch/issue510. Can you have a look, 
please? Fortunately, it's a rather small patch.

Instead of

size_t num_operators = operators.size();
    for (size_t i = 0; i < num_operators; ++i)
        do_something(operators[i]);

we'd like to write code à la

for (Operator op : operators)
    do_something(op);