Hardware-Efficient Neural Networks for Low-Latency Multiplexed Superconducting Qubit Readout | Synapse