The Inference-Compute Frontier and a Latency-Efficient Architecture for Limit Order Book Prediction