Technically, because we don't actually switchroot into this filesystem
we don't need some of the directories and symlinks (as they are already
present in the initrd).

But with a long term view to making this a "real" filesystem that we DO
switchroot into, it makes sense to try and keep things vaguely valid.

This allows us to actually mount and run the system inside a container for
more advanced debugging.

I've successfully booted my systemd, logged in, had a fiddle and then
powered it off again. There are some delays and it's requried to umount
the /mnt/proc which seems to get automatically mounted before running
nspawn. Also audit support has to be turned off in the kernel too to
allow a full boot to work properly (otherwise the authentication in
the container doesn't work. This is a known problem in the kernel and
will be fixed eventually.

This commit converts the rescue system to systemd rather than the legacy
sysvinit.

Besides being faster, one primary advantage of running systemd here is
that we can "boot" the system we're inspecting via systemd-nspawn
to properly test it.

Note: Bits that don't work:
 * nspawn
 * old dracut initqueue udev rules in /etc/ are not cleaned out (drop in
   support error?)
 * the gui may loop differently now