phrases and AND queries are currently resolved by
multi-merging the vectors of the files involved. This
could (probably) be more efficiently performed using a
heap. Test to make sure it is actually faster, or at
least doesn't degrade as badly as i think the current
one will :o(