Fix __int128_t redefinition in int128.h !CONFIG_INT128 fallback on Clang/GCC

[robertream]

Problem

Building QEMU's int128.h fails on Clang/GCC whenever the build reaches the
!CONFIG_INT128 fallback:

qemu/include/qemu/int128.h:150:16: error: typedef redefinition with different types
    ('Int128' (aka 'struct Int128') vs '__int128')
typedef Int128 __int128_t;

The fallback aliases the reserved builtin name __int128_t to struct Int128
so fallback code can name the type. That is only valid on compilers that do not
provide __int128_t as a builtin. On GCC/Clang the builtin always exists (when
128-bit ints are supported), so the alias is a redefinition error.

This bit two real setups:

• Apple clang on arm64 (macOS), and
• the Rust unicorn-engine-sys crate's vendored build (its cmake path ends up
  without CONFIG_INT128, so it compiles the fallback — and fails on any modern
  clang/gcc).

#2251 added the current guard #if !(defined(_MSC_VER) && defined(__clang__)),
but that only covers clang-cl; regular Clang/GCC still hit the redefinition.

Fix

Gate the alias on !defined(__SIZEOF_INT128__). __SIZEOF_INT128__ is the standard
macro that is defined iff the compiler provides the 128-bit builtin, so:

• GCC/Clang (incl. clang-cl): builtin present → alias skipped (uses the builtin) ✅
• MSVC: no builtin → alias emitted (as before) ✅

This subsumes #2251 and fixes the general case.

Validation

• tests/regress/int128_redefinition.c — a self-contained reproduction of the
  fallback construct. It fails to compile with the old guard on any
  __int128-capable compiler and compiles cleanly with the fix (verified on Apple
  clang / arm64, producing the exact error above pre-fix).
• .github/workflows/int128-regression.yml — compiles + runs that test on
  macos-14 (Apple Silicon, arm64 clang), macos-13 (Intel) and Linux.

Why a new job: existing CI never exercises this fallback on arm64 clang — the
C-library CI builds on macos-14 via cmake with CONFIG_INT128 defined (native
path, fallback not compiled), and the Rust-crate CI (Crate-publishing.yml) does
not run on Apple Silicon.

Happy to drop the test/CI or fold it into an existing workflow if maintainers prefer.

[wtdcode]

Hi please note our PR should go against the dev branch.

Also tests go to tests/unit.