fix: gate AVX/AVX-512/AMX dispatch on OS XSAVE support via XGETBV

[Mattbusel]

Problem

ggml_backend_cpu_x86_score() has carried a // FIXME: this does not check for OS support comment since this file was first written.

The function gates SIMD dispatch purely on CPUID feature flags — but CPUID only reports hardware capability. The OS must also have opted in to saving and restoring the relevant extended register state in XCR0. On systems where it hasn't (certain hypervisors, containers, custom kernels), executing a VEX-encoded (AVX/AVX2) or EVEX-encoded (AVX-512) instruction causes an Illegal Instruction fault (#UD / SIGILL) at runtime, even though CPUID says the CPU supports it.

This is the root cause behind a class of bug reports such as:

• Compilation / distribution of a whisper.dll that 'just works' and is performant across all x86 processors, whether they support AVX2 extensions or not #1939 — "Whisper.dll fails on x86 processors without AVX2"
• Incredibly slow on Windows with CPU having AVX support #630 — "Incredibly slow on Windows with CPU having AVX support"
• Suboptimal performance in SSE-enabled CPUs that don't have AVX #251 — "Suboptimal performance in SSE-enabled CPUs that don't have AVX"

The same fix was recently applied to the standalone ggml-org/ggml repo (from which this file is vendored).

Fix

Add three helpers to cpuid_x86:

Helper	XCR0 bits	Gates
os_saves_ymm()	[2:1] — SSE + YMM	AVX, AVX2, FMA, F16C, AVX-VNNI
os_saves_zmm()	[7:5] — opmask + ZMM_Hi256 + Hi16_ZMM	All AVX-512 variants
os_saves_amx()	[18:17] — XTILECFG + XTILEDATA	AMX-INT8

xgetbv() is portable: _xgetbv() intrinsic on MSVC, inline xgetbv asm on GCC/Clang — matching the same pattern already used for cpuid/cpuidex in this struct.

Impact

• No behavioural change on any correctly configured system (XCR0 is always set when the CPU advertises the feature under a normal OS).
• Prevents SIGILL under hypervisors or containers that do not set XCR0 to match CPUID.
• Removes the FIXME that has existed since this code was written.