Reduce LOOKUP JOIN memory usage

[richardkchapman]

In cases where the number of input rows can be known ahead of time, it would be advantageous for both fragmentation and performance reasons to pre-allocate the hash table and has rows as the are received, rather than first gathering into an expanding array and then hashing into a second array.

Also, as currently coded, both the 'gather array' and the hash table are kept in separate halves of the same allocated array - this means that the gather array cannot be released once the hash table is built.

[ghalliday]

Thor should also use packed for the rows allocated for the rhs of the join. This is a special case of a more general issue that is included somewhere else (but I couldn't find.)