Instead of stopping the service, running migrations etc (those take
time), and then starting again, let's only restart the service after the
uprada has already happened. Now that we are using unicorn, which
supports restarting with dropping connections, with this we should have
0 downtime upgrades.

Also removed some unecessary buts deom debian/noosfero.prerm:

- remove restarting of memcached, which is broken behavior as it could
  affect other services using memcached. Just set a namespace for our
  cache keys, and all caches will be automatically invalidated when the
  application is restarted after an upgrade with a new version number

- remove removal of cache/*, since script/production already does that.

NOTE: since this behavior change depends on changes to the prerm
script, this won't take effect until the next upgrade _after_ the
upgrade to first version including this change.