In SELECT_LEX, unify fields_list (visible fields) and all_fields
(visible _and_ invisible fields); before this patch, they were
overlapping linked lists (except in a few cases where they were
not, undocumented). This is necessary but not sufficient to be
able to have visible and invisible items in arbitrary order;
in order to get that, we will also need to remove most cases of
thd->change_item_tree(), and a few other things, detailed in
comments. However, this is a good cleanup in itself, and enables
us to use e.g. range-based for loops in a number of places.

The new data structure is an std::deque of Item *, which has
much less confusing semantics, is more cache-efficient and also
works without constant pointer chasing. (Ideally, we'd move
towards a Mem_root_array, but cannot do it before the aforementioned
changes.) This type change propagates far into the code, but is mostly
localized to the optimizer. (It removes about 80% of all instances of
List<Item>, which is a good step on the way to removing List<T> entirely.)
Item has received a boolean that signifies whether it is hidden or not,
and many loops that iterated over fields_list now check this flag and
skip hidden ones.

As List used empty() to clear a list, but std::deque<T> uses that
to return whether the list is empty or not (so fields.empty() is
a no-op with no compiler warnings), there was a huge risk of
introducing wrong code during merges and development, ie., code
that would call fields.empty(), causing very hard-to-find
bugs. For this reason, List<T>::empty() has been renamed to
List<T>::clear(). is_empty() has _not_ been renamed to empty(),
so that we get compile errors on such merges and can catch them
early.

Many places that sent a non-const reference to List<Item> have
been changed to either a const reference or a pointer, as the
style guide mandates (most of this code predates the style guide).
This makes it clear whether the fields list is to be modified
or not.

Change-Id: I520f680416f5bf2dcdd61d002eea02878cb1dc10