Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuning