opportunistic cmpxchg on set insert

We now "guess" that the target slot we're inserting into is empty and perform a
compare-exchange expecting this to be the case. This works out because, if it is
not empty, we just performed the read we needed to anyway. It doesn't seem like
this is much of a performance win, but it makes the code slightly more elegant
so I've committed it anyway.

Closes Github #31 "opportunistic cmpxchg"