io assumptions

Just a warning: the store IO open and create callers assume that if a
non null pointer is returned, then the object will not have called it's
close callback yet. This is a) potentially false given a really weird
scheduling event and something like overlapped IO. And b) duplication of
effort (for async opens, it means we will fail if we can't queue the
request, or if the open fails - and the caller is given the
responsiblity to handle that difference. A cleaner design IMO would be a
single failure mode - we queue the open in store_swapin, and if the open
fails (sync or async) we get called back. If the open succeeds, we get a
separate callback with the newly created object.

This is a note for any interested hacker, not something I will be
addressing anytime soon.

Rob