Parallelize some steps of experiment swapin. Mainly, we now fork off
another process to handle os_setup, and let that run while doing things like snmpit. This should make experiment creation slightly faster, and hide some of the scaling time of some scripts (particularly, snmpit) behind the much larger node reboot time. Things that _have_ to be done before the nodes come up (namely exports_setup and starting the event system) are done before the os_setup process is forked, just to be absolutely sure that they get done before they're needed.