Magnitude Invariant Parametrizations Improve Hypernetwork Learning