Learnability with PAC Semantics for Multi-agent Beliefs