Use per-consumer/provider locks from mtx_pool() to serialize the statistics
update instead of global ones. Additional lock acquisition for queued
dispatch may add some overhead, but instead it removes any collisions for
direct dispatch.