[Feature request] Use a stable GGML CPU thread pool instead of re-creating it for each request

**Motivation:** Right now, it appears that each GGML graph computation gets a new thread pool on the CPU path. We ran some benchmarks, and using a stable thread pool led to a ~5% performance improvement on Android. Our hypothesis is that this is due to the OS having better statistics on the thread usage and placing them on the proper cores. We already have a thread-count heuristic that also tries to accomplish this by matching the # of threads to the # of high-performance cores, but we still get a small boost with a stable thread pool. We avoided direct thread-pinning to give the OS freedom to make the best decision based on the current device state.