Rework the way we turn on and off performance schema batch mode in iterators.
The signal to turn on and off is now exclusively sent through iterators,
instead of looking into QEP_TABs and cleaning up directly on handlers.

In particular, this means that we send the “end PFS batch mode” signal to
every single iterator in the tree at the end, guaranteeing that no involved
table is left behind. This is a bit more brute-forcish than trying to turn
it off on only the ones where we've turned it on, but much, much simpler to
get right.

The method for starting batch mode is also changed; instead of trying to
figure out from the outside whether NestedLoopIterator is involved or not,
we call StartPSIBatchMode() on the root iterator, which will send the signal
down to the table iterator if we have a single table; if not, it will be
stopped and ignored at the first NestedLoopIterator. Again, this vastly
simplifies the logic.

Change-Id: I8ad0e11e0fc0db6a54174c1b841c103f57fab32d