Resolving Discrepancies in Compute-Optimal Scaling of Language Models