What Matters More For In-Context Learning under Matched Compute Budgets: Pretraining on Natural Text or Incorporating Targeted Synthetic Examples?