Fixing the NTK: From Neural Network Linearizations to Exact Convex Programs