Before the cost of an aggregate distinct (COUNT(DISTINCT ...)) was set
to 0 if the values where part of an index and the cost of grouping
was higher than the best cost so far.  This was shown in explain with
"Using index for group-by (scanning)".

This patch fixes it by calculating the cost of aggregate distinct
and using scanning only if the cost was better than group-by-optimization.

Thing taken into account:
- When using aggregate distinct on index, the filtering is done before
  the row is checked against the WHERE and we have thus less WHERE cost.
- When comparing a cost from aggregate distinct, we add to the compared
  to plan the cost of doing the filtering later in the SQL level.