Reported by Wolfgang Richter, Jul 1, 2010

As per design discussions, allow the in memory hashtable to fill to 
a certain configurable/overridable value (passed through the 
properties of a job).

Flush to a file when it fills, merging if the file already exists.

Reducers just merge all data together (in the same way as a mapper). 
and once they have everything from all mappers (get the all finished 
signal), they write out the final results to the DFS.

Also, both mappers and reducers should have a final merge even if 
their hashtable isn't full whenever they finish processing all of 
their input data.