It depends on your purpose. If you have lots of samples (e.g. >100) and want to call SNPs in these samples, multi-sample calling is preferred as it produces low false positive rate (FPR). Errors will be added up if you call each sample separately and then combine the calls. Even 1% FPR per sample will be greatly amplified given 100 samples. You may see lots of spurious singletons. On the other hand, if you want to get the best consensus for each sample, you should call each sample separately. Joint calling has higher false negative rate for each sample. No single strategy suits all purposes.