Anatomy of a high-performance EP kernel

How expert-parallel dispatch and combine kernels work, built up from
scratch: the high-throughput shape and the low-latency one.