Free Bits: Latency Optimization of Mixed-Precision Quantized Neural Networks on the Edge