Nice bite-sized commits:

• Improve formatting of _Default_allocate_traits; define a new macro instead of conditionally inserting _CONSTEXPR20. Drive-by: Combine nested conditionals.

• Fix TRANSITION comment for disabled EDG config in P1502R1_standard_library_header_units.

• Annotate workaround for Clang constrained-friends issue with TRANSITION, Clang 19. I discovered this has been fixed in 19.1.0 after reducing a repro to file, didn't try to dig up the fix.

• Remove extraneous constexpr in P2502R2_generator

• Simplify unreachable_sentinel_t. We had to workaround the fact that hidden friends aren't hidden in permissive mode by stuffing unreachable_sentinel_t's operator== into a detail namespace. Now that the changes to C++20 relational expressions are implemented, we can use a member operator== instead of a friend to more simply achieve the same effect. See https://godbolt.org/z/x5ffjdxYf for a proof-of-concept.

• Uniformly annotate undefined members of fake iterators

• Cleanup some iterator definitions

  • Remove nested type names that are extraneous in C++20 to help us avoid depending on the existence of said nested type names.
  • Always declare the "nested 5 iterator traits" (or the subset of them a given iterator provides) in canonical order at the top of the first public section in a class body.
  • When the "nested 5 iterator traits" are available, use their names in the semantically appropriate places in member function declarations. For example, reference operator*() const.

• Rename _Merge_move to _Merge_move_unchecked for consistency with other algorithms that take unwrapped iterators.

• Improve detection of library-defined array specializations. _Unchecked_begin/_Unchecked_end are semi-documented, so their presence doesn't necessarily indicate a specialization is library-defined. Use our _Is_from_primary tech, and remove the now-unused _Has_unchecked_begin_end.

• Add some internal checks to internal algorithms. _Uninitialized_copy, _Uninitialized_copy_n, and _Uninitialized_move in <xmemory> are internal-only, and trust that the caller has passed arguments that denote a valid range. I am not so trusting. _STL_INTERNAL_CHECK((_STD _Adl_verify_range(_First, _Last), true)) feels like a bit of a hack - I could be convinced to simply wrap the _Adl_verify_range call in #ifdef _ENABLE_STL_INTERNAL_CHECK.

• Verify that arguments to subrange's iter/sentinel constructors denote a range. They must per N4988 [range.subrange.ctor] para 1 and 3. As a rule, we never verify that [begin(meow), end(meow)) for a range meow actually denotes a range, so subrange's constructors are the only chance we have to verify these user-supplied values.