Not-Just-Scaling Laws: Towards a Better Understanding of the Downstream Impact of Language Model Design Decisions