Use 4.7x fewer synchronization barriers in GGML

This commit and the previous one provide a big win for CPUs like EPYC
and Threadripper which have lots of cores. I'm now getting 839.57 tok
per second on Mistral 7b v0.2 bf16 prefill, which is much better than
the 557 tok/sec mentioned in the blog post. Generating speed has been
increased ~15-20% on my machine compared to our last release.