llama.cpp

threads: improve ggml_barrier scaling with large number of threads

#9598

Merged