Computational Tradeoffs of Optimization-Based Bound Tightening in ReLU Networks