Passing the size directly lowers the number of operations performed by
bidirectional stable_partition in the general case (collection not
already partitioned or mostly partitioned). This in turn improves the
speed of bidirectional slabsort over random data.

This avoids a full collection scan when slab_sorter is given a full
std::list to sort.

With bidirectional iterators it potentially avoids n operations needed
to compute the size of the sequence.

The internal traits were accidentally guarded by __cpp_lib_range since
version 1.9.0, because I put the #endif in the wrong place.

This avoids accidentally specifying usig a different iterator tag for a
sorter's iterator_category and its internal static_assert, leading to
fewer potential bugs.

As a result, slabsort could only work with iterators for which an
ADL-found iter_swap was available.

Fixed/tweaked the following things about the class group_iterator used
in the implementation of merge_insertion_sort:
- Remove non-const operator[] (iterators aren't supposed to have one)
- Ensure it can wrap forward iterators
- Kill warning in make_group_iterator when difference_type < int
- Reduce type & code indirections where they are unneeded