Make counting examined_rows the responsibility of the relevant iterators,
instead of the join machinery (as the old join machinery will not be used in
the iterator executor).

An alternative would be counting number of rows that come out and number of rows
that are somehow filtered away at various stages, but it rapidly becomes very
complicated and fiddly to get right.

This also fixes an issue where temporary tables that were sorted got some rows
counted three times in examined_rows instead of twice as they should be (once
for read from the temporary table, once for read from the sort buffer).

Filesort optimizer trace now no longer returns the number of unfiltered records,
since WL #12074 is going to pull filtering out of filesort altogether. Instead,
"num_found_rows" reports consistently the number of rows processed by filesort,
whether we are using priority queue or not. If there's no condition to filter on,
this will be exactly the same as "num_examined_rows" used to print.

Change-Id: Ief1168f640313b08e3b0c115ad19cf9525f77b6c