Asymptotics of Learning with Deep Structured (Random) Features