Edge Inference with Fully Differentiable Quantized Mixed Precision Neural Networks