FIXME this can't go into blead as is

FIXME this can't go into blead as is

Optimization to make the Schwartzian transform faster

This patch consists of two conceptually separate but technically
entwined changes. In a nutshell, the idea is to detect the use of an
array dereference for sorting on precomputed data. This means sort
invocations like the following:

  sort { $a->[0] <=> $b->[0] }

and

  sort { $a->[0] cmp $b->[0] }

as well as the reversed cases.

The two major changes are:

1) Detection of a basic Schwartzian transform-style sort blocks
and removal of the corresponding bits of OP tree. Sets a new flag on
the sort OP which will indicate that this sort needs to be performed
by one of the special-case optimized C functions in pp_sort.

Right now, we only match array dereferencing using the constant 0.
Technically, it is trivial to match any other constant, and feasible to
match the use of a variable. But I could not think of a good way to pass
the array index to pp_sort.

2) Implementation of special-purpose sort functions that do
array-dereferencing. This change implements the above sort functions for
all of numeric, integer, and string sorts, as well as overloaded values
in the arrays and locale-string sorts.

It supports overload magic inner SVs (see tests added in next
commit). I believe it also handles the locale cmp case in the same way
as the normal optimized sort functions do, but since I don't properly
understand locales, the tests don't cover that.

The net effect of this is a consistent speed-up by over 2x on my
machine when doing a Schwartzian transform or similar technique
involving sorting on the first element of an array. I tried a variety
of array sizes, as well as inclusion of the two map stages of the
idiom.