terminated; this makes the final assert in the self-test code fail if
the parent runs faster than the children. Fix this by calling wait()
on the remaining children instead.
os.fdopen() calls unbuffered. I presume that it's enough if we can
make all three of them (for stdin, stdout, and stderr) unbuffered and
don't need to specify different buffer sizes per file -- that would
complicate the interface more than I care for.
