I guess we were causing the kernel to syslog on every single packet on
MacOS.  Oops.

If the subprocess was trying to write to its stdout/stderr, its process
would never actually finish because it was blocked waiting for us to read
it, but we were blocked on waitpid().  Instead, use waitpid(WNOHANG) and
continually read from the subprocess (which should be a blocking operation)
until it exits.

Not sure if this will fix anything, but it might stop the problem reported
on some MacOS versions where the firewall doesn't get cleaned up correctly.

(apenwarr: also updates to the matching, latest minimal/do)

This makes it easier (possible?) to generate sshuttle.8 from sshuttle.md on
MacOS.  We also import the git-enhanced version numbering magic so the
generated manpage can have a real version number.

Now that we imported the feature from redo, might as well use it.

On top of the bug that already existed in 10.6, Lion also makes the sysctl
needed to fix the problem into a read-only variable, so we have to actually
change it at kernel boot time and force people to reboot.  Nice job, Apple.

And make sshuttle exit with a well-defined exit code (111) if it needs to
reboot.

If the expected arch directory doesn't exist, give up and don't specify arch at
all. Currently it expands to '*' which fails.

[slightly modified by apenwarr]