regression from commit 63d2a603b9b2bc27090405ba4a05a4aaa8840ea2

rationale: we drop support in stage2 and their support was totally
broken due to missing conversion to dracut, udev, ...

Now that dracut runs stage1 earlier, we need to store it's exit code
for later reuse. In this, the separate dracut module will just write a
flag file in /run/drakx/ to indicate that we want to re-exec /sbin/init
(after doing some symlinks and /usr bind mount) rather than run the install.

As a result there is no need to return 0x35 ('procced') so just return
0 and keep 0x35 return code for when we really do need to restart stage1
(e.g. on a fatal error). The dracut module should loop if 0x35 is
ever returned.

init is now solely responsibe for running stage2 (be it install or rescue)
as stage1 is now run earlier by dracut.

Dracut mounts these already.

In order to run stage1 we now need to do things a little differently,
namely that we move the symlinking fixups into 'init' itself which we will
actually call as a kind of fake switch_root implementation to avoid
hacking dracut too much[1]. As we run stage1 during dracut's 'mount' phase
we should not do the symlink hacks so early otherwise it will mess up
the rest of the dracut initqueue processing.

Currently, the initrd's /usr is nuked and then symlinked, but to
allow for easier debugging I now just bind mount the stage2 /usr over
the top which allows us to unmount it again easily enough if we want to
go back.

We no longer touch the /bin, /sbin or /lib[64] symlinks as these are
relative and simply bind mounting /usr is enough to fix all of them.

1. This scheme will have to change if we eventually swtich to a
systemd based dracut invocation.