Pull filtering out of filesort; now that we have iterators, there's no need
for filesort to worry about checking a condition or unlocking rows.
Instead, we can just put a FilterIterator in front of it.

There's still one quirk here, in that filesort discards the read sets and
builds up new ones from scratch. For this, it still needs to have access
to the condition, to know which fields it might need. This is a layering
violation and stops us from being able to sort the output of a join directly
without the need for a temporary table, but it will remain for the time being.

Change-Id: Iea7e97e56b9479fb470e3e580160e89e953d90f3