ggml-cpu: skip unsupported ARM ISA variants under GGML_CPU_ALL_VARIANTS

[ninihuang2026]

Overview

When building with GGML_CPU_ALL_VARIANTS, probe the exact -march= string each ARM variant would be built with and skip it if the compiler rejects it, instead of failing the whole build, because not every compiler version supports every ARM ISA extension (e.g. GCC < 14 rejects "+sme").

Additional information

1. Issue links:

• Fixes Compile bug: zero-size array ‘gemm_gemv_kernels’ / invalid feature modifier ‘sme’ #14464 : This resolves the open bug report
• Compile bug: invalid feature modifier ‘sme’, when building on ARM with GGML_CPU_ALL_VARIANTS=ON #17403: The issue was closed via a GCC 14 workaround. In there, user @o7si posted a suggestion in the comment that wraps armv9.2_1 and armv9.2_2 in a single shared if(COMPILER_SUPPORTS_SME) check, one probe, using only armv9.2_1's flag combo, gating both variants. The proposed approach in this PR probes each ARM variant independently via its own check_cxx_compiler_flag() call.

2. Evidence
   Test env:
   Jetson Orin (Cortex-A78AE, Armv8.2-A, aarch64), Ubuntu 24.04.4 LTS, GCC 13.3.0 (Ubuntu 13.3.0-6ubuntu2~24.04.1)

commands:

cmake -S . -B build -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON
cmake --build build -j$(nproc)

===before===
cc1plus: error: invalid feature modifier 'sme' in '-march=armv9.2-a+dotprod+fp16+sve+i8mm+sme'
Build exits non-zero, entire GGML_CPU_ALL_VARIANTS build aborts.

===after===

-- Performing Test GGML_CPU_VARIANT_armv9.2_1_SUPPORTED - Failed
-- Skipping CPU backend variant 'armv9.2_1': compiler does not support '-march=armv9.2-a+dotprod+fp16+sve+i8mm+sme'
-- Performing Test GGML_CPU_VARIANT_armv9.2_2_SUPPORTED - Failed
-- Skipping CPU backend variant 'armv9.2_2': compiler does not support '-march=armv9.2-a+dotprod+fp16+sve+i8mm+sve2+sme'

The other 6 ARM variants (armv8.0_1 through armv8.6_2) still probe Success and build with their correct differentiated -march= flags
The full build completes (exit 0) — llama-cli, libllama.so, libggml.so, plus the 6
ggml-cpu-armv8.*.so backend variants.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. Claude Sonnet suggested the patch and tested and verified the result.