NB this is so we can put things in LOGFILE at earlier points
than the script does at present

We capture the error code at a few points in the script, and then (intend to)
log something about what went wrong. But if we set -e we immediately bail
before logging anything useful!

Resolves #1

We now e.g. have a cron job on cerebro-backup which calls
these scripts, where /sbin is not on the $PATH.

For cerebro-backup where we do backups by the directory rather than the
whole server. We need to have one machine being backed up
(cerebro-filestore) but a task per directory.

The new-backup-rsnapshot script understands a 'postgres' argument, but
this set up a postgres backup in an old style that we no longer use.
This change updates it to do some of the work of setting up a new style
postgres backup and tell the user what else they might need to edit to
make it go; it varies quite a lot depending on server.

From the comment in the script, I think this was an attempt to ignore an
error about setting GIDs. The only place I saw that particular error
occuring in practice was for loopback-mounted isos on OS builders.
We shouldn't be backing those up in the first place! Exit code
23 is "partial transfer due to error" and so can be due to problems
other than failing to set GIDs

Split logs for tasks with same target machine but different target filesystems into
different logfiles for easier debugging.

prepare-nondebian script creates a fake excludes file for the machine now if
one doesn't exist, because I keep forgetting to do it myself for Redhat
machines and then the zfs-rsync script fails.

Forcing arcfour causes us not to be able to back up jessie machines.