Physics-Informed Design of Input Convex Neural Networks for Consistency Optimal Transport Flow Matching