Beyond Chinchilla-Optimal: Accounting for Inference in Language Model Scaling Laws