Bug #30921780: ROLLUP OUTPUT IS NOT REASONABLE WHEN SAME FIELDS EXIST IN THE GROUP BY SYNTAX
Bug #26227613: REPEATING COLUMN IN GROUP BY WITH ROLLUP YIELDS WRONG RESULT
Bug #29134467: REFACTOR RECURSION IN RESOLVE_ROLLUP_ITEM TO USE ITEM_TRANSFORM
Bug #30967158: THE ROLLUP SUPER-AGGREGATE ROW OUTPUT IS NOT CORRECT

Reimplement rollup without using any slices at all. This is both an important
step on the way towards removing the slice system entirely, and a simplification
in its own right.

Instead of using slices and refs, we introduce special items during resolving
that operate differently depending on which rollup level we are on; there is one
for group items and one for sums. (The AggregateIterator has a list of relevant
Items, and signals them when switching rollup level.) The biggest challenge is
keeping these correct during creation of temporary tables; the techniques we
introduce (mostly various walks down the Item trees) may very well be important
to remove slices in other places, too. There are some rollup items that are
left in the lower temporary tables where they could have been removed, but they
are generally harmless.

The concepts of overlapping arrays of sum_funcs (with sum_func_end) is no
longer needed and has been removed. Similarly, the special item for rollup
constants is gone. copy_sum_funcs() was only ever used by AggregateIterator,
and has been moved into a private member function.

Since we now explicitly resolve fields in the SELECT list to group items,
and have a central place for doing so (SELECT_LEX::find_in_group_list),
it is easy to change the resolver rules. We use this to change the resolving
of aliases to prioritize matches to group items to the same alias, to give
slightly more logical behavior when there are repeated group items; there are
illegal queries as per the SQL standard, and we explicitly make no guarantee
as of future behavior, but we now match what 5.7 did, after 8.0.11 changed
the behavior.

Change-Id: I3485d18f3f0671f9dcda8509322f313b4b10307a