DEPENDENCY INTO ACCOUNT

Earlier, the following algorithm was used to order tables 
before doing greedy search: 
 1) Outer tables before inner tables 
    (as in "outer LEFT JOIN inner")
 2) Increasing number of expected rows 

By taking key dependency into account when doing this 
pre-sorting, greedy search will be able to considerably 
reduce the number of plans because pruning is more likely
to happen.

In addition, merge and insert sort is implemented because
the std::stable_sort implementation differs between platforms.
This results in different QEPs which in turn is a problem for 
mtr testing.