Pipes should be created with O_CLOEXEC

On Linux all file handles are inherited by default by child processes. This means that when you create an `opstream` and attach it to a child process, the child process with inherit the input side of the pipe. When you try to close the pipe from the parent process, the operating system will still consider the input of the pipe to be open in the child and the process will most likely block indefinitely. This can be worked around by passing a `boost::process::pipe` created from the result of `pipe2(pipefd, O_CLOEXEC)`, however it would be better to have this behaviour by default. Please note that this must be done at the initial `pipe` call and not with `fcntl` later as otherwise there would be a race condition in multithreaded code.