Fix: GCC Target Mismatch for vdotq_s32 on AArch64 (Apple M1, M2, etc)

[russellwmy]

Problem

Building llama.cpp on AArch64 (e.g. in Docker on macOS M1/M2) fails when compiling NEON dot product code using vdotq_s32. GCC throws the following error:

error: inlining failed in call to 'always_inline' 'int32x4_t vdotq_s32(...)': target specific option mismatch

This error occurs because vdotq_s32 requires the ARMv8.x-A dot product extension (+dotprod), which is not enabled by default in most toolchains.

This PR disables GGML's use of native optimizations as a temporary fix, without modifying llama.cpp directly. Since Apple Silicon CPUs may not be a primary target for the llama.cpp project, this workaround allows successful builds.

Feel free to close this PR if it does not apply to your use case.

[MarcusDunn]

I think this triggers too often (i.e. some builds on arm64 linux will want native optimizations). I'm not opposed to allowing this to be disabled, but an env var or a disabled default feature would be preferred.