debugging: guard ABSL_HAVE_ELF_MEM_IMAGE on <link.h> availability

[vasko-110]

Fixes #2153.

ABSL_HAVE_ELF_MEM_IMAGE is enabled for every __ELF__ target except a deny-list, and the header then includes <link.h> unconditionally. __ELF__ describes the object format, not the presence of a dynamic loader, so bare-metal ELF toolchains (arm-none-eabi with newlib) fail with a fatal error. This tests for the header the code actually needs instead.

The existing exclusions are untouched: __sun, __QNX__ and __OpenBSD__ do ship a <link.h> with different semantics, so they must stay listed regardless.

Verified on 71330b93: elf_mem_image.cc and vdso_support.cc now compile under arm-none-eabi-g++, and on the host ABSL_HAVE_ELF_MEM_IMAGE is still defined and both files still compile — no supported platform changes behaviour.

[mkruskal-google]

Instead of just disabling this logic for bare-metal, we could add support like we did for FreeBSD. Something like this seems like it should work, if we also condition the include of link.h:

#ifndef ElfW
#  if defined(__ELF_NATIVE_CLASS)
     /* FreeBSD native infrastructure uses __ElfN */
#    define ElfW(type) __ElfN(type)
#  elif defined(__x86_64__) || defined(__aarch64__) || defined(__riscv) && (__riscv_xlen == 64)
     /* Fallback block for 64-bit bare-metal environments */
#    define ElfW(type) Elf64_##type
#  else
     /* Fallback block for 32-bit bare-metal environments */
#    define ElfW(type) Elf32_##type
#  endif
#endif```

[vasko-110]

You're right, it builds — newlib ships <elf.h> even though it has no <link.h>, so this is enough:

#if __has_include(<link.h>)
#include <link.h>  // for ElfW
#else
#include <elf.h>
#endif

#ifndef ElfW
#if defined(__FreeBSD__)
#define ElfW(type) __ElfN(type)
#elif __SIZEOF_POINTER__ == 8
#define ElfW(type) Elf64_##type
#else
#define ElfW(type) Elf32_##type
#endif
#endif

Two changes from yours. __ELF_NATIVE_CLASS is glibc's, not FreeBSD's — glibc's link.h defines ElfW before we get here, so that branch never fires, and FreeBSD would quietly lose __ElfN. And __SIZEOF_POINTER__ saves listing ppc64/s390x/mips64/loongarch64.

I'd still rather keep it off here, though. VDSOSupport::Init() needs either getauxval or /proc/self/auxv, so IsPresent() is always false on bare metal — that's 5 KB of .text and open/read/close in the link for a lookup that can't succeed.

Either way works for me, which do you prefer?