Revert r285968: Always use parallel_for_each because it falls back to std::for_each.

It turned ou that we actually want to call std::for_each even if
threading is supported. Unless --thread is given, LLD shouldn't use
more than one threads.

llvm-svn: 286004