Principled Weight Initialisation for Input-Convex Neural Networks