Enforcing balance allows local supervised learning in spiking recurrent networks