The exact technical explanation is, for any Machine Learning model you need training data. The more the training data, the better the performance. Given huge number of Tesla's in California (and Tesla running its own calibration vehicles), there is more training data to train the FSD models. Hence the performance in California is better.
To give even more technical explanation, in Machine Learning models you need both global features (general road and traffic sign shapes) and local features (specific type of road in a specific area on a highway). In California, given more Tesla drivers (and Tesla running its own calibration vehicles), they have been able to collect more local features. In other areas with less Teslas, they have data on less local features. So the performance is better in California.