Add a benchmark for precompleted context creation