construct the hashtables iteratively to save memory up to B times

[xu3kev]

Since each hashtable can be constructed once and then discarded after the union-find step, we can rearrange the for loop into constructing the B hashtables one by one instead of constructing them all at once. (Note that each hashtable consumes a large amount of memory because it holds the (document idx, hash) for the entire dataset.) We can then potentially save the memory usage by up to B times.