Run the last iteration of parallel_for_loop using a threadpool.

Remainders of tasks were ran in the main thread, so parallel_for_each
could theoretically take 2x time than the ideal.

llvm-svn: 288631