i686: libstdc++ 15 headers fail with "__gnu_cxx has no member __bfloat16_t" (missing __BFLT16_* predefined macros) #511

In the edg_x86_32 test configuration (the gcc-i686 base, which uses the
system libstdc++ 15 headers with -m32), any test that includes <meta> (or
another header that pulls in <type_traits>, <cmath>, <ostream>, etc.)
fails to compile in C++23/C++26 mode:

"/usr/include/c++/15/type_traits", line 514: error: namespace "__gnu_cxx" has no member "__bfloat16_t"
      struct __is_floating_point_helper<__gnu_cxx::__bfloat16_t>
                                                   ^

"/usr/include/c++/15/bits/cpp_type_traits.h", line 348: error: namespace "__gnu_cxx" has no member "__bfloat16_t"
      struct __is_floating<__gnu_cxx::__bfloat16_t>
                                      ^
...
8 errors detected in the compilation of "Test_name.c".

So far this has been papered over by recording the failure:
reflections/demo1 has a .demo1.rto/edg_x86_32.1.1.txt recording and a
FAIL(COMPILE) entry in tests/expectations/edg_x86_32/reflections.log, and
the new reflection tests in #241 need the same treatment. That means the
reflection tests effectively don't run on i686.

Cause

libstdc++ uses and defines the type under different conditions:

• The specializations above are guarded by #ifdef __STDCPP_BFLOAT16_T__, which
  the front end predefines whenever extended floating-point types are enabled
  (src/macro.c).

• The typedef itself, in bits/c++config.h (both the 64-bit one and the
  32/ multilib one), is guarded by __BFLT16_DIG__:

  #if defined __cplusplus && defined __BFLT16_DIG__
  namespace __gnu_cxx
  {
    typedef __decltype(0.0bf16) __bfloat16_t;
  }
  #endif

bases/docker/dev-env/gcc/lib_linux_x86_64/predefined_macros.txt defines
__BFLT16_DIG__ (and the other __BFLT16_* and __FLT16_* macros) in GNU
mode, but lib_linux_i686/predefined_macros.txt (identical in the gcc and
gcc-i686 bases, unchanged since the initial release) has none of them. Yet
GCC 15 predefines them with -m32 too:

$ echo | g++ -m32 -std=c++23 -dM -E -x c++ - | grep -E 'BFLT16_DIG|FLT16_DIG'
#define __BFLT16_DIG__ 2
#define __FLT16_DIG__ 3

Possible fix

Add the __BFLT16_* and __FLT16_* entries that GCC predefines with -m32 to
lib_linux_i686/predefined_macros.txt, mirroring the gnu entries in the
x86-64 file. Then re-record the affected edg_x86_32 recordings
(reflections/demo1 and the tests from #241) and drop their FAIL entries
from tests/expectations/edg_x86_32/.

Also worth checking: plain g++ -m32 (no SSE2) does not define
__STDCPP_BFLOAT16_T__, while the front end always does when extended
floating-point types are enabled. With __BFLT16_DIG__ defined, libstdc++'s
uses become consistent again, so that difference may not matter in practice.