Unpacking Tokenization: Evaluating Text Compression and its Correlation with Model Performance