fix crash on non-AVX systems dynamically loading GGML CPU backends

[jmorganca]

Thanks for the awesome work by @slaren in #10469 (and a few follow up PRs) to enable dynamic GGML backend loading. This made supporting different CPU instructions in GGML much, much easier.

I noticed a small hitch with with the llamafile code where a machine with a non-AVX CPU would crash when trying to dlopen CPU libraries built with GGML_LLAMAFILE=ON. This moves the AVX-dependent code to do a member variable, fixing the crash on dlopen. I'm not sure how sgemm.cpp is vendored, and so let me know the best way/place to suggest a change.

[slaren]

Thanks, I missed this global. The fix looks ok, but if the code is not inlined it may add some overhead to the other types. I will leave this open for a while in case someone knowledgeable about llamafile/tinyblas wants to propose a better solution.

[jmorganca]

Thanks for merging @slaren. I'm running some performance tests after noticing ollama/ollama#9087. I'm not sure if this PR is the root cause, but I haven't ruled it out yet. In any case will keep you posted and wanted to give you a heads up in case

[slaren]

Llamafile tinyblas should only be used for prompt processing, so if you are also observing a decrease of performance during generation, it is not very likely that it was caused by this change.