fix: guard llama.cpp backend init against SIGILL on x86_64 CPUs without FMA/AVX2

[mvanhorn]

Summary

On Linux x86_64, goosed aborts with SIGILL on older Intel CPUs that lack FMA (the reporter's Ivy Bridge i5-3230M has AVX + F16C but no FMA/AVX2). The faulting instruction is vfmadd213ss inside ggml_cpu_init, which the bundled llama.cpp (ggml) CPU backend runs unconditionally when LlamaBackend::init() is called from LlamaCppBackend::new(). Because the illegal instruction takes down the whole backend process, the Desktop frontend then loses every backend-dependent feature (Skills, Apps, session history, ACP websocket).

This adds an x86_64 CPU capability precheck that runs before LlamaBackend::init(). When a required instruction set is missing, it returns a clear, actionable error naming the missing features instead of letting ggml execute an illegal instruction, so local inference is cleanly disabled and the rest of the backend stays alive.

• check_cpu_supports_local_inference() is gated by #[cfg(target_arch = "x86_64")] and uses std::arch::is_x86_feature_detected! to check the instruction sets the bundled ggml CPU build relies on: FMA, AVX2, F16C, BMI2, and SSE4.2.
• On non-x86_64 targets (ARM, macOS metal) it compiles to an unconditional Ok(()), leaving those paths untouched.
• The error message names every missing instruction set so users get an actionable diagnostic.

Testing

cargo test -p goose --features local-inference --lib -- providers::local_inference::llamacpp (the two added unit tests pass):

• local_inference_cpu_support_check_accepts_current_host derives its expectation from the host's detected features, so it asserts Ok on a supported host and Err on an x86_64 host missing the required sets (the exact environment this fix handles) rather than spuriously panicking there.
• local_inference_cpu_support_error_names_missing_instruction_sets checks the error message names the missing sets.

Also verified with cargo fmt --check and cargo clippy -p goose --features local-inference -- -D warnings.

Related Issues

Closes #10073

Screenshots/Demos (for UX changes)

N/A. This is a backend change to local inference initialization with no UI surface.

[jbg]

As a follow-up I'll look into building llama.cpp with wider CPU support; it already has runtime feature detection so we should be able to support local inference on CPUs without FMA/AVX2. This fix is solid in the meantime though.

[mvanhorn]

Thanks for merging, @jbg - guarding llama.cpp init against SIGILL keeps older x86_64 CPUs from hard-crashing.