Federated Deep Equilibrium Learning: A Compact Shared Representation for Edge Communication Efficiency